diff --git a/QUICKSTART.md b/QUICKSTART.md
index 5d5f8ab78..f2ab57f63 100644
--- a/QUICKSTART.md
+++ b/QUICKSTART.md
@@ -81,8 +81,9 @@ works.
 
 ## 3. Start the server
 
-Check your file descriptor limit — EverOS opens many LanceDB segment
-files under concurrent search + indexing. Platform defaults:
+With the default LanceDB index backend, check your file descriptor limit —
+EverOS opens many LanceDB segment files under concurrent search + indexing.
+Platform defaults:
 **macOS 256** · **Linux 1024** · **Windows 8192**. If yours is below
 4096, raise it before starting:
 
@@ -258,7 +259,8 @@ This is what makes EverOS different — memory persists as plain Markdown:
 ├── ome.toml                            ← strategy config (hot-reloaded)
 ├── .index/                             ← derived indexes (rebuildable from md)
 │   ├── sqlite/system.db
-│   └── lancedb/
+│   ├── lancedb/                        ← default derived index backend
+│   └── milvus/                         ← Milvus Lite DB when configured
 └── .tmp/
 ```
 
diff --git a/README.md b/README.md
index f423d467a..905fc8e9d 100644
--- a/README.md
+++ b/README.md
@@ -39,8 +39,10 @@
 EverOS is a Python library and local-first memory runtime for agents and
 makers. It gives one portable memory layer across coding assistants, apps,
 devices, and workflows from day one. It stores conversations, files, and agent
-trajectories as readable Markdown, then syncs local SQLite and LanceDB indexes
-for fast retrieval and self-evolving reuse.
+trajectories as readable Markdown, then syncs local SQLite and a derived
+vector/BM25 index for fast retrieval and self-evolving reuse. LanceDB is the
+default embedded index backend; Milvus can be enabled when you want the same
+rebuildable index to run on Milvus Lite, Milvus server, or Zilliz Cloud.
 
 <table>
 <tr>
@@ -60,7 +62,7 @@ for fast retrieval and self-evolving reuse.
 </tr>
 <tr>
 <td><strong>Local three-part stack</strong></td>
-<td>✅ Markdown + SQLite + LanceDB; no MongoDB, Elasticsearch, or Redis required</td>
+<td>✅ Markdown + SQLite + embedded LanceDB by default; optional Milvus when configured</td>
 <td>❌ Often depends on managed services, vector DBs, graph DBs, or server stacks</td>
 </tr>
 <tr>
@@ -114,6 +116,12 @@ uv pip install everos
 # or: pip install everos
 ```
 
+To use Milvus as the derived index backend:
+
+```bash
+uv pip install "everos[milvus]"
+```
+
 ### 2. Play With The Demo
 
 Run this before configuring API keys or starting the server:
@@ -642,7 +650,7 @@ Explore stored entities and relationships in a graph interface. Frontend demo; b
 ## Documentation
 
 - [docs/everos-demo.md](docs/everos-demo.md) — Demo scope and TUI source layout
-- [docs/how-memory-works.md](docs/how-memory-works.md) — Markdown, SQLite, LanceDB, and recall flow
+- [docs/how-memory-works.md](docs/how-memory-works.md) — Markdown, SQLite, derived index backends, and recall flow
 - [docs/use-cases.md](docs/use-cases.md) — Full use-case gallery and integration examples
 - [docs/engineering.md](docs/engineering.md) — Contributor engineering reference: build, test, CI, conventions
 - [docs/migration-to-1.0.0.md](docs/migration-to-1.0.0.md) — Legacy API migration notes
diff --git a/config.example.toml b/config.example.toml
index f8043127f..0145dd686 100644
--- a/config.example.toml
+++ b/config.example.toml
@@ -56,3 +56,19 @@ max_concurrent  = 5
 #
 # [lancedb]
 # read_consistency_seconds = 5.0
+
+# ── Optional Milvus derived index ─────────────────────
+# Markdown remains the source of truth. Set the index backend to Milvus
+# when you want the rebuildable vector/BM25 index to live in Milvus.
+# Leave uri empty for Milvus Lite at <root>/.index/milvus/milvus.db.
+#
+# [index]
+# backend = "milvus"
+#
+# [milvus]
+# uri = ""                         # or "http://localhost:19530"
+# token = ""                       # required for Zilliz Cloud/auth-enabled Milvus
+# db_name = ""
+# consistency_level = "Session"
+# dimension = 1024                 # must match your embedding model output
+# collection_prefix = "everos"
diff --git a/docs/api.md b/docs/api.md
index b76ca99fb..19e68d803 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -84,8 +84,9 @@ bare FastAPI `detail`); see [Errors](#errors).
 
 `/add` and `/flush` write the markdown file (the source of truth)
 **synchronously** — when the call returns with `status: "extracted"`,
-the new entry exists on disk. The LanceDB vector / BM25 / scalar index
-is rebuilt by the in-process **cascade coroutine asynchronously**.
+the new entry exists on disk. The configured vector / BM25 / scalar
+index backend is rebuilt by the in-process **cascade coroutine
+asynchronously**.
 
 That means `/search` and `/get` may not see a record immediately after
 the `/flush` that produced it. Typical sync latency is sub-second, but
@@ -362,7 +363,7 @@ A recursive boolean tree of predicates. Used by `/search.filters` and
 `/get.filters`. The Pydantic envelope only checks the recursive
 combinator shape; field-level validity (which scalar fields are
 filterable, which operators apply, value coercion) runs when the
-node is compiled to a LanceDB `where` clause server-side. Compile
+node is compiled to a backend-specific filter clause server-side. Compile
 errors surface as `422` with the offending field / operator in
 `error.message`.
 
@@ -455,12 +456,12 @@ Examples:
 |---|---|
 | `"keyword"` | BM25 only — pure lexical match, no embedding cost |
 | `"vector"` | Dense vector ANN only — semantic recall, no lexical |
-| `"hybrid"` *(default)* | Reciprocal-rank fuse of BM25 + vector + optional scalar filter in a single LanceDB query |
+| `"hybrid"` *(default)* | Reciprocal-rank fuse of BM25 + vector + optional scalar filter against the configured derived index backend |
 | `"agentic"` | Iterative cluster-path retrieval driven by a cross-encoder rerank loop; higher quality at higher latency / cost |
 
 `"hybrid"` is the default because it balances recall and precision
-with one LanceDB roundtrip. `"agentic"` calls the LLM in a loop and
-should be reserved for offline or background workflows.
+without requiring the agentic loop. `"agentic"` calls the LLM in a loop
+and should be reserved for offline or background workflows.
 
 ### GetMemoryType
 
@@ -596,7 +597,7 @@ scope.
   this `(session_id, app_id, project_id)`, or it was already flushed).
 
 `/flush` is synchronous with respect to markdown persistence: by the
-time the response returns, the new entry is on disk. LanceDB index
+time the response returns, the new entry is on disk. Derived index
 sync is still asynchronous — see
 [Eventual consistency](#eventual-consistency).
 
diff --git a/docs/architecture.md b/docs/architecture.md
index afb5efe30..4aae7826a 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -17,7 +17,7 @@
 │    + reflection + strategies + get + events            │
 ├──────────────────────────────────────────────────────┤
 │  infra/persistence  (Storage adapters; infra/ may host other adapter types)    │
-│    markdown + sqlite + lancedb                        │
+│    markdown + sqlite + derived index                  │
 └──────────────────────────────────────────────────────┘
 
 Cross-cutting (used by all layers, depends on none):
@@ -68,8 +68,8 @@ layers = [
 └────────────────────────────────────────────────────────────────┘
 
    ┌──────────────┐   ┌──────────────┐   ┌─────────────────┐
-   │   Markdown   │   │   SQLite     │   │    LanceDB      │
-   │  (truth)     │   │  (state)     │   │  (index)        │
+   │   Markdown   │   │   SQLite     │   │ Derived index   │
+   │  (truth)     │   │  (state)     │   │ LanceDB/Milvus  │
    ├──────────────┤   ├──────────────┤   ├─────────────────┤
    │ entries +    │   │ change queue │   │ vector ANN      │
    │ frontmatter  │   │ + state/LSN  │   │ BM25 (Tantivy)  │
@@ -78,7 +78,7 @@ layers = [
    └──────────────┘   └──────────────┘   └─────────────────┘
           │                  │                    │
           ▼                  ▼                    ▼
-    memory-root/         .index/sqlite/      .index/lancedb/
+    memory-root/         .index/sqlite/      .index/<backend>/
    (truth source)       (system data)       (rebuildable)
 ```
 
@@ -101,10 +101,13 @@ External message
    │        │
    ▼        ▼
 4a. SQLite   4b. memory.cascade  (async daemon)
-    audit        watches md → diff entries → LanceDB sync
+    audit        watches md → diff entries → index sync
 ```
 
-**Key guarantee**: md write is strongly consistent (fsync). LanceDB is eventually consistent. LanceDB unavailability does not block response — changes buffer in the SQLite `md_change_state` queue, replayed on recovery.
+**Key guarantee**: md write is strongly consistent (fsync). The derived
+index is eventually consistent. Index backend unavailability does not block
+response — changes buffer in the SQLite `md_change_state` queue, replayed on
+recovery.
 
 ## Read path
 
@@ -115,8 +118,8 @@ User query
 1. service.search
    │
    ▼
-2. memory.search (hybrid)     single LanceDB query =
-                                BM25 + vector ANN + scalar filter
+2. memory.search (hybrid)     BM25 + vector ANN + scalar filter
+                              through the configured index backend
    │
    ▼
 3. (optional) read md         original markdown for context
@@ -139,12 +142,13 @@ extract/
 
 ### `memory/cascade/`
 
-Daemon that watches markdown changes and syncs to LanceDB:
+Daemon that watches markdown changes and syncs to the configured derived
+index backend:
 
 - inotify / FSEvents file watcher (cross-platform via `watchdog`)
 - 500ms debounce
 - Entry-level diff (added / changed / removed)
-- LanceDB single-transaction update (text + vector columns atomic)
+- Per-entry index upsert / delete (text + vector columns update together)
 - LSN-based crash recovery via the SQLite `md_change_state` queue
 - Handlers for all eight business kinds: episode, atomic_fact, foresight,
   user_profile, agent_case, agent_skill, knowledge_document, knowledge_topic
@@ -225,7 +229,7 @@ holding **only memory extraction algorithms**:
 everalgo is:
 
 - **Stateless** — pure functions, no class hierarchy
-- **No I/O** — does not touch md files / LanceDB / SQLite
+- **No I/O** — does not touch md files, derived indexes, or SQLite
 - **No prompts inline** — extractors that accept a prompt-override parameter use the project-supplied value; others use their algo-bundled defaults
 
 This boundary lets everalgo be reused across product forms (this open-source build, EverOS Cloud, OpenClaw plugins, etc.).
diff --git a/docs/cascade_runbook.md b/docs/cascade_runbook.md
index c9811eeea..d7bb9d530 100644
--- a/docs/cascade_runbook.md
+++ b/docs/cascade_runbook.md
@@ -1,9 +1,13 @@
 # Cascade Runbook
 
-The cascade daemon keeps LanceDB in sync with the markdown files under
-the memory root. Service / entry points only ever write markdown; the
-daemon is the **sole** writer of the LanceDB index. This runbook covers
-the recurring operational questions.
+The cascade daemon keeps the configured derived index in sync with the
+markdown files under the memory root. Service / entry points only ever write
+markdown; the daemon is the **sole** writer of the derived index. This runbook
+covers the recurring operational questions.
+
+Sections that mention LanceDB-specific schemas, index cache, file descriptors,
+or `lance error` messages apply to the default LanceDB backend. Milvus uses the
+same cascade queue with backend-specific collection management.
 
 ## What runs where
 
@@ -13,7 +17,7 @@ providers in order:
 1. **Metrics** — Prometheus collector.
 2. **LLM** — LLM client initialisation.
 3. **SQLite** — system DB + schema (`SQLModel.metadata.create_all`).
-4. **LanceDB** — async connection + schema verification + FTS indexes.
+4. **Derived index** — async connection + schema verification + search indexes.
 5. **Cascade** — watcher + scanner + worker, all in-process tasks.
 6. **OME** — offline memory engine.
 
@@ -23,7 +27,7 @@ The cascade subsystem itself is three independent loops:
 |---|---|---|
 | Watcher | `watchdog` filesystem events (sync thread) | `md_change_state.upsert` per registered kind |
 | Scanner | Periodic walk (`scan_interval_seconds`, default 30 s) | Same — catches changes the watcher missed |
-| Worker | `claim_pending_batch` polling (default 1 s when idle) | Handler dispatch → LanceDB upsert / delete |
+| Worker | `claim_pending_batch` polling (default 1 s when idle) | Handler dispatch → index upsert / delete |
 
 Every loop talks to the same `md_change_state` sqlite table. The
 worker's claim mode (`pending → processing → done/failed`) keeps
diff --git a/docs/cli.md b/docs/cli.md
index e24e6e5ea..c56574ada 100644
--- a/docs/cli.md
+++ b/docs/cli.md
@@ -3,7 +3,7 @@
 The `everos` command-line entry point covers **setup and operations** —
 generate starter config files (`init`), run the HTTP API server (`server
 start`), inspect effective config (`config show`), and operate the
-md → LanceDB index queue (`cascade`). Hot-path
+md → derived index queue (`cascade`). Hot-path
 business (`/add` `/flush` `/search` `/get`) is the **HTTP API**, not the
 CLI.
 
@@ -32,7 +32,7 @@ everos
 │   └── show [--root PATH]          Show effective configuration
 ├── server
 │   └── start [--host] [--port] [--root] [--reload] [--log-level]   Start the HTTP API server (uvicorn)
-└── cascade [--root PATH]           Inspect / operate the md → LanceDB sync queue
+└── cascade [--root PATH]           Inspect / operate the md → derived index sync queue
     ├── status                      Queue / LSN summary
     ├── sync [PATH]                 Drain the queue now (optional PATH force-enqueues)
     └── fix [--apply]               List failed rows / re-enqueue retryable ones
@@ -43,8 +43,9 @@ Each subcommand lives in its own module under
 registered in `cli/main.py`. The CLI is intentionally small — hot-path
 business (`/add` `/flush` `/search` `/get`) is the **HTTP API**, not the
 CLI; the CLI covers setup (`init`), running the server, and index ops
-(`cascade`). There is no `reindex` command — rebuild by deleting
-`<root>/.index/lancedb` and restarting, or run `everos cascade sync`.
+(`cascade`). There is no `reindex` command — rebuild by deleting the
+selected backend's derived index directory (`<root>/.index/lancedb` or
+`<root>/.index/milvus`) and restarting, or run `everos cascade sync`.
 
 ## `everos server start`
 
@@ -68,8 +69,8 @@ everos server start \
 | `--root` | `EVEROS_ROOT` | `~/.everos` |
 | `--reload` | — | off (use in development) |
 
-Lifespan startup wires the storage backends (SQLite engine + LanceDB
-connection) on app boot; see
+Lifespan startup wires the storage backends (SQLite engine + configured
+derived index backend) on app boot; see
 [`entrypoints/api/lifespans/`](../src/everos/entrypoints/api/lifespans/).
 
 ## Configuration via env vars
@@ -81,7 +82,9 @@ Both CLI and HTTP server read configuration from `pydantic-settings`:
 | `EVEROS_ROOT` | memory-root path (default `~/.everos`) |
 | `EVEROS_MEMORY__TIMEZONE` | `Settings.memory.timezone` (e.g. `Asia/Shanghai`) |
 | `EVEROS_SQLITE__BUSY_TIMEOUT_MS` | `Settings.sqlite.busy_timeout_ms` |
+| `EVEROS_INDEX__BACKEND` | `Settings.index.backend` (`lancedb` or `milvus`) |
 | `EVEROS_LANCEDB__READ_CONSISTENCY_SECONDS` | `Settings.lancedb.read_consistency_seconds` |
+| `EVEROS_MILVUS__URI` | `Settings.milvus.uri` |
 
 Pattern: `EVEROS_<SECTION>__<KEY>` (double underscore = nesting). See
 [`config/settings.py`](../src/everos/config/settings.py).
diff --git a/docs/configuration.md b/docs/configuration.md
index 98fdf3479..6848d7839 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -99,6 +99,26 @@ everos init --root /data/everos
 | `read_consistency_seconds` | float \| null | `null` | Read consistency interval. `null` = no check, `0` = strict, `>0` = eventual. |
 | `index_cache_size_bytes` | int | `16777216` | Upper bound on LanceDB index cache (16 MB default). |
 
+### `[index]`
+
+| Field | Type | Default | Description |
+|---|---|---|---|
+| `backend` | `"lancedb"` \| `"milvus"` | `"lancedb"` | Rebuildable vector/BM25 index backend used by cascade, search, and get. Markdown remains the source of truth; SQLite remains the system state store. |
+
+### `[milvus]`
+
+Milvus is optional. Install with `everos[milvus]`, then set
+`[index].backend = "milvus"`.
+
+| Field | Type | Default | Description |
+|---|---|---|---|
+| `uri` | string \| null | `""` | Empty uses Milvus Lite at `<root>/.index/milvus/milvus.db`. Set to a Milvus server or Zilliz Cloud endpoint to use an external service. |
+| `token` | string \| null | `""` | Token for Zilliz Cloud or auth-enabled Milvus. |
+| `db_name` | string \| null | `""` | Optional Milvus database name. |
+| `consistency_level` | `"Strong"` \| `"Session"` \| `"Bounded"` \| `"Eventually"` | `"Session"` | Milvus consistency level for created collections. |
+| `dimension` | int | `1024` | Dense vector dimension. Must match the configured embedding model output. |
+| `collection_prefix` | string | `"everos"` | Prefix for EverOS Milvus collections. |
+
 ### `[llm]`
 
 | Field | Type | Default | Required | Description |
@@ -228,3 +248,5 @@ Examples:
 | `[llm] api_key = "sk-..."` | `EVEROS_LLM__API_KEY=sk-...` |
 | `[sqlite] busy_timeout_ms = 10000` | `EVEROS_SQLITE__BUSY_TIMEOUT_MS=10000` |
 | `[memory] timezone = "Asia/Tokyo"` | `EVEROS_MEMORY__TIMEZONE=Asia/Tokyo` |
+| `[index] backend = "milvus"` | `EVEROS_INDEX__BACKEND=milvus` |
+| `[milvus] uri = "http://localhost:19530"` | `EVEROS_MILVUS__URI=http://localhost:19530` |
diff --git a/docs/how-memory-works.md b/docs/how-memory-works.md
index 1a8136478..b3d25ba1d 100644
--- a/docs/how-memory-works.md
+++ b/docs/how-memory-works.md
@@ -23,14 +23,16 @@ This is the narrative companion to the reference docs: see
 
 ## The storage stack
 
-Three embedded pieces, each owning what it is best at. Markdown is the
-**source of truth**; the other two are **derived and rebuildable**.
+Three pieces, each owning what it is best at. Markdown is the
+**source of truth**; the other layers are **derived and rebuildable**.
+By default the vector/BM25 index is embedded LanceDB. When configured,
+Milvus can serve the same derived index role.
 
 | Layer | Backed by | Holds | Rebuildable? |
 |---|---|---|---|
 | **Markdown + YAML frontmatter** | plain `.md` files | the memory content itself — the only portable, human-editable asset | — (it *is* the truth) |
 | **SQLite** (`aiosqlite`) | `.index/sqlite/*.db` | system state, audit log, the cascade queue, the boundary buffer, OME engine state | ✅ from markdown |
-| **LanceDB** (Arrow) | `.index/lancedb/*.lance` | vector + BM25 + scalar columns for retrieval | ✅ from markdown |
+| **Derived index** | `.index/lancedb/*.lance` by default, or `.index/milvus/milvus.db` for Milvus Lite | vector + BM25 + scalar columns for retrieval | ✅ from markdown |
 
 !!! note "The one rule that follows from this"
     Delete the entire `.index/` directory and **no memory is lost** — it
@@ -77,8 +79,10 @@ visually distinct from a user-named one).
 │   │   ├── ome.db                          Offline Memory Engine state
 │   │   ├── ome.aps.db                      APScheduler jobstore (split to avoid lock contention)
 │   │   └── ome.db.lock                     OME single-engine guard (portalocker)
-│   └── lancedb/
-│       └── <kind>.lance/                   one Arrow table per kind
+│   ├── lancedb/
+│   │   └── <kind>.lance/                   default derived index backend
+│   └── milvus/
+│       └── milvus.db                       Milvus Lite when configured
 │
 ├── ome.toml                                ← user-editable OME strategy overrides (hot-reloaded)
 └── .tmp/                                   atomic-write staging
@@ -96,7 +100,7 @@ visually distinct from a user-named one).
 The path manager is
 [`MemoryRoot`](../src/everos/core/persistence/memory_root.py); every path
 above is a property on it. `MemoryRoot.ensure()` creates the runtime dirs
-(`.index/{sqlite,lancedb}/`, `.tmp/`); user-visible dirs appear on first
+(`.index/{sqlite,lancedb,milvus}/`, `.tmp/`); user-visible dirs appear on first
 write. Config files (`everos.toml`, `ome.toml`) are created by `everos init`.
 
 ## How a memory is born
@@ -129,7 +133,7 @@ index catches up asynchronously.
                                                            ▼
                                        md_change_state queue (SQLite, durable)
                                                            ▼
-                                            rebuild LanceDB rows  ──▶  searchable
+                                            rebuild index rows  ──▶  searchable
 ```
 
 - **`/add`** appends messages to a per-`(session_id, app_id, project_id)`
@@ -142,7 +146,7 @@ index catches up asynchronously.
 - Everything else (atomic facts, foresight, profile, agent cases/skills)
   is produced **asynchronously** by the OME — see
   [the OME section](#the-offline-memory-engine-ome).
-- The **cascade daemon** turns every `.md` write into LanceDB rows so the
+- The **cascade daemon** turns every `.md` write into derived index rows so the
   content becomes searchable.
 
 ## Memory types & storage strategies
@@ -177,7 +181,8 @@ The three strategies:
 
 ## The cascade daemon
 
-The cascade subsystem keeps LanceDB in sync with the markdown tree. It runs
+The cascade subsystem keeps the configured derived index in sync with the
+markdown tree. It runs
 **in-process** with the server (a coroutine started by the app lifespan),
 not as a separate OS daemon.
 
@@ -187,7 +192,7 @@ not as a separate OS daemon.
    durable, so a crash mid-sync replays on restart.
 3. A worker drains the queue at **entry-level** granularity: it diffs the
    file, re-embeds only changed entries (keyed by `content_sha256`), and
-   upserts the LanceDB rows.
+   upserts the configured index rows.
 
 Because markdown is the source of truth, **editing a file directly is
 fully supported** — open an episode in VSCode / Obsidian / Vim, change an
@@ -257,8 +262,8 @@ Two paths, two guarantees:
 
 | Path | Guarantee | Detail |
 |---|---|---|
-| **Write** (`/add`, `/flush`) | **strong** | the episode `.md` is on disk before the call returns `extracted`; never blocks on LanceDB |
-| **Read** (`/search`, `/get`) | **eventual** | reads LanceDB, which lags md by the cascade processing time — sub-second typically, up to ~10–15 s under load |
+| **Write** (`/add`, `/flush`) | **strong** | the episode `.md` is on disk before the call returns `extracted`; never blocks on the derived index |
+| **Read** (`/search`, `/get`) | **eventual** | reads the configured derived index, which lags md by the cascade processing time — sub-second typically, up to ~10–15 s under load |
 
 So a `/search` immediately after the `/flush` that produced a record may
 miss it. The markdown is durable regardless; index lag never loses data. If
@@ -272,17 +277,21 @@ entry needs re-embedding; an LSN watermark (in `system.db`) orders
 rebuilds; the durable `md_change_state` queue is the replayable audit
 trail.
 
-## Zero external services
+## Zero external services by default
 
 No database server, message broker, or vector service to run. Vector ANN,
-full-text BM25, and scalar filtering all execute inside the **embedded
-LanceDB** engine in one query; SQLite is a local file. The whole stack is a
-single directory you can copy, back up, or check the user-visible parts of
-into git.
+full-text BM25, and scalar filtering execute inside the embedded LanceDB
+backend by default; SQLite is a local file. The whole default stack is a single
+directory you can copy, back up, or check the user-visible parts of into git.
+
+If you set `[index].backend = "milvus"`, EverOS uses Milvus as the same
+rebuildable derived index. Empty `[milvus].uri` uses Milvus Lite under the
+memory root; setting `uri` / `token` points the index at Milvus server or
+Zilliz Cloud.
 
 !!! note
     There is no automatic "grep over markdown" search fallback today — if
-    the LanceDB index is unavailable, rebuild it from markdown (it is
+    the configured index is unavailable, rebuild it from markdown (it is
     derived and disposable) rather than relying on a degraded search path.
 
 ## Operating it
@@ -294,12 +303,12 @@ The CLI ([cli.md](cli.md)) is intentionally small:
 | `everos init` | generate starter config files (`everos.toml` + `ome.toml`) |
 | `everos server start` | run the HTTP API (cascade + OME start with it) |
 | `everos cascade status` | queue / LSN summary |
-| `everos cascade sync` | drain the cascade queue now (force md → LanceDB) |
+| `everos cascade sync` | drain the cascade queue now (force md → derived index) |
 | `everos cascade fix` | list failed rows / re-enqueue retryable ones |
 
 !!! warning "There is no `everos reindex` or `everos flush`"
     - **Reindex** = the index is rebuildable: stop the server,
-      `rm -rf <memory-root>/.index/lancedb`, restart — the cascade
+      `rm -rf <memory-root>/.index/lancedb` (or `.index/milvus` for Milvus Lite), restart — the cascade
       rebuilds from markdown. For an incremental catch-up, use
       `everos cascade sync`.
     - **Flush** is an HTTP endpoint (`POST /api/v1/memory/flush`), not a
diff --git a/docs/knowledge.md b/docs/knowledge.md
index c2aa81698..11f40ca70 100644
--- a/docs/knowledge.md
+++ b/docs/knowledge.md
@@ -48,8 +48,8 @@ Each level corresponds to a different granularity of API:
 ## Storage layout
 
 Every document is a self-contained directory. Markdown files are the
-single source of truth; SQLite and LanceDB are derived indexes built
-automatically by the cascade daemon.
+single source of truth; SQLite and the configured vector/BM25 index are
+derived indexes built automatically by the cascade daemon.
 
 ```
 ~/.everos/<app>/<project>/knowledge/
@@ -73,16 +73,16 @@ automatically by the cascade daemon.
 ### Storage roles
 
 ```
-Markdown (source of truth)  +  SQLite (structured state)  +  LanceDB (vector + BM25 index)
+Markdown (source of truth)  +  SQLite (structured state)  +  derived vector/BM25 index
 ```
 
 | Store | What it holds | Role |
 |-------|---------------|------|
 | Markdown | Document metadata, summaries, topic content, original files | Single source of truth; human-readable and editable |
 | SQLite | Document rows, topic rows (with content), change queue | Structured queries, paginated lists, count aggregation |
-| LanceDB | Topic vectors, BM25 tokens, scalar fields | Search index (fully rebuildable from Markdown) |
+| Derived index | Topic vectors, BM25 tokens, scalar fields | Search index (fully rebuildable from Markdown) |
 
-Even if SQLite and LanceDB data is corrupted, as long as the Markdown
+Even if SQLite and derived index data is corrupted, as long as the Markdown
 files are intact, the indexes can be fully rebuilt via the cascade daemon.
 
 ### Markdown format
@@ -500,7 +500,7 @@ query ─→ embed ─→ keyword (BM25) ─┐
 ```
 
 1. **Embed** — the query is embedded using the configured embedding provider
-2. **Recall** — dual-channel retrieval from LanceDB:
+2. **Recall** — dual-channel retrieval from the configured derived index:
    - BM25 channel: keyword matching on `summary_tokens` + `content_tokens`
    - ANN channel: nearest-neighbor search on the `vector` column
    - In `hybrid` mode, both channels run in parallel
@@ -531,7 +531,7 @@ export EVEROS_KNOWLEDGE__SEARCH__RERANK_N=100
 ## Cascade sync
 
 The cascade daemon watches the knowledge Markdown directory for file
-changes and keeps SQLite + LanceDB in sync.
+changes and keeps SQLite + the configured derived index in sync.
 
 ```
 md file written
@@ -539,7 +539,7 @@ md file written
   → worker picks up from queue (≤1s poll interval)
   → handler dispatched by file type:
       index.md    → KnowledgeDocumentHandler → SQLite upsert (metadata)
-      N_topic.md  → KnowledgeTopicHandler    → tokenize + embed + SQLite + LanceDB upsert
+      N_topic.md  → KnowledgeTopicHandler    → tokenize + embed + SQLite + index upsert
 ```
 
 The topic handler uses a SHA-256 content digest to skip unchanged files —
@@ -597,7 +597,7 @@ curl .../documents?app_id=tenant_b&project_id=proj_1
 # → { "documents": [], "total": 0 }
 ```
 
-Storage paths, SQLite rows, and LanceDB indexes are all scoped by
+Storage paths, SQLite rows, and derived index rows are all scoped by
 `app_id` + `project_id`.
 
 ## End-to-end walkthrough
diff --git a/docs/multimodal.md b/docs/multimodal.md
index 0dc9e4ce3..e4fdcdfb4 100644
--- a/docs/multimodal.md
+++ b/docs/multimodal.md
@@ -41,7 +41,7 @@ POST /api/v1/memory/add
   boundary detector → extraction LLM → memory cell (MemCell)
         │
         ▼
-  markdown (truth)  +  SQLite (state)  +  LanceDB (vector + BM25)
+  markdown (truth)  +  SQLite (state)  +  derived vector/BM25 index
         │
         ▼
   retrievable via /search and /get like any text memory
diff --git a/docs/overview.md b/docs/overview.md
index 7f9be38ec..dccf68120 100644
--- a/docs/overview.md
+++ b/docs/overview.md
@@ -11,7 +11,7 @@ Build an open-source Python memory framework where **AI agents' long-term memory
 - Local deployment for personal agents or small teams
 - Conversation, workflow, agent-trace, file-knowledge → structured memory
 - Hybrid retrieval (BM25 + vector + scalar filter)
-- Cascade index sync (md edit → LanceDB sub-second)
+- Cascade index sync (md edit → derived index)
 - Dual-track memory (user-track / agent-track)
 - Offline memory evolution (Foresight / AtomicFact / Profile / Skill),
   including Reflection — a consolidation strategy within the OME that
@@ -30,7 +30,7 @@ Build an open-source Python memory framework where **AI agents' long-term memory
 ### 1. Markdown as Source of Truth
 
 ```
-delete all LanceDB / SQLite files → can rebuild from md
+delete all derived index / SQLite files → can rebuild from md
 delete any md file               → memory is gone
 ```
 
@@ -42,7 +42,7 @@ User trust comes from physical visibility — the user can `cat` / `vim` / `grep
 |---|---|---|
 | Markdown files | Truth source — entries, frontmatter | Search (grep is degraded fallback only) |
 | SQLite | Queue, cascade audit log, sensitive data isolation | Vector / full-text |
-| LanceDB | Vector ANN + BM25 + scalar filter, single-query hybrid | Be the source of truth (loss = rebuild from md) |
+| Derived index | Vector ANN + BM25 + scalar filter | Be the source of truth (loss = rebuild from md) |
 
 ### 3. Algorithm-orchestration separation
 
diff --git a/docs/reflection.md b/docs/reflection.md
index 94a56aec3..620fcddd2 100644
--- a/docs/reflection.md
+++ b/docs/reflection.md
@@ -147,14 +147,15 @@ narrative.
 
 ## Storage layout
 
-Memory uses Markdown as the single source of truth; SQLite and LanceDB are
-derived indexes built automatically by the cascade daemon.
+Memory uses Markdown as the single source of truth; SQLite and the configured
+vector/BM25 index backend are derived indexes built automatically by the
+cascade daemon.
 
 | Store | What it holds | Role |
 |---|---|---|
 | Markdown | Episode bodies, merged narratives, archive markers | Single source of truth; human-readable and editable |
 | SQLite | Clusters and members, consolidation audit records | Structured state and queries |
-| LanceDB | Vectors + BM25 index for Episodes / atomic facts | Search (rebuildable from Markdown) |
+| Derived index | Vectors + BM25 index for Episodes / atomic facts | Search (rebuildable from Markdown) |
 
 The **merged narrative** is written to the Episode daily-log Markdown; its
 frontmatter marks that it came from a cluster:
@@ -186,7 +187,7 @@ deprecated_entries:
 ---
 ```
 
-> Soft-archive, not delete: even if SQLite / LanceDB are corrupted, as long
+> Soft-archive, not delete: even if SQLite / the derived index are corrupted, as long
 > as the Markdown is intact the indexes can be fully rebuilt — and every
 > consolidation remains traceable back to its original content.
 
@@ -355,5 +356,5 @@ curl -s -X POST "$BASE/memory/search" \
 ## See also
 
 - [how-memory-works.md](how-memory-works.md) — Episodes and the memory extraction pipeline
-- [storage_layout.md](storage_layout.md) — Markdown + SQLite + LanceDB stack
+- [storage_layout.md](storage_layout.md) — Markdown + SQLite + derived index stack
 - [api.md](api.md) — full HTTP API reference
diff --git a/docs/storage_layout.md b/docs/storage_layout.md
index 860446648..71447905a 100644
--- a/docs/storage_layout.md
+++ b/docs/storage_layout.md
@@ -3,8 +3,9 @@
 How `everos` lays out a memory-root on disk: directory tree, file
 naming, frontmatter chassis, and entry-id encoding.
 
-The contents are the **source of truth**; SQLite and LanceDB are
-derived indexes that can be rebuilt from markdown alone.
+The contents are the **source of truth**; SQLite and the configured
+vector/BM25 index backend are derived indexes that can be rebuilt from
+markdown alone.
 
 ## 1. Memory-root tree
 
@@ -49,8 +50,10 @@ the frontmatter (see [§3](#3-frontmatter-chassis-yaml)).
 │   │   ├── ome.db                        Offline Memory Engine state
 │   │   ├── ome.aps.db                    APScheduler jobstore (split to avoid lock contention)
 │   │   └── ome.db.lock                   OME single-engine guard (portalocker)
-│   └── lancedb/
-│       └── <kind>.lance/                one directory per LanceDB table
+│   ├── lancedb/
+│   │   └── <kind>.lance/                default derived index backend
+│   └── milvus/
+│       └── milvus.db                    Milvus Lite database when configured
 │
 ├── ome.toml                             user-editable OME strategy overrides (hot-reloaded)
 └── .tmp/                                staging dir for batch / multi-step writes
@@ -64,7 +67,7 @@ the frontmatter (see [§3](#3-frontmatter-chassis-yaml)).
 
 The path manager is [`MemoryRoot`](../src/everos/core/persistence/memory_root.py),
 exposing every path as a property. `MemoryRoot.ensure()` creates the
-runtime-required dirs (`.index/{sqlite,lancedb}/`, `.tmp/`); the
+runtime-required dirs (`.index/{sqlite,lancedb,milvus}/`, `.tmp/`); the
 user-visible dirs are *not* pre-created — they appear on first write.
 Config files (`everos.toml`, `ome.toml`) are created by `everos init`.
 
@@ -168,9 +171,9 @@ Implementation: [`core/persistence/markdown/entries.py`](../src/everos/core/pers
 > **File-level seq, not global**: the same `ep_20260601_00000001` may
 > appear across two different `user_id`s (each user has its own daily file).
 > Cross-table joins must therefore key on **`(scope_id, entry_id)`**
-> rather than `entry_id` alone — see SQLite/LanceDB tables that follow.
+> rather than `entry_id` alone — see the derived index tables that follow.
 
-## 5. SQLite + LanceDB derived indexes
+## 5. SQLite + derived indexes
 
 ```
 .index/
@@ -178,9 +181,10 @@ Implementation: [`core/persistence/markdown/entries.py`](../src/everos/core/pers
 │   └── system.db          state / audit / cascade queue + buffer / LSN
 │                           (system tables: md_change_state, memcell,
 │                            unprocessed_buffer, conversation_status, cluster)
-└── lancedb/
-    └── <kind>.lance/      one Arrow table per business kind — the per-kind
-                            rows (text / vector / tokens / metadata) live here
+├── lancedb/
+│   └── <kind>.lance/      default derived index backend
+└── milvus/
+    └── milvus.db          Milvus Lite database when configured
 ```
 
 - **SQLite** ([`infra/persistence/sqlite/tables/`](../src/everos/infra/persistence/sqlite/tables/))
@@ -190,13 +194,14 @@ Implementation: [`core/persistence/markdown/entries.py`](../src/everos/core/pers
   per-kind business rows. `reflection_report` is the audit trail for
   Reflection merges (cluster_id, mode, source_members, merged_entry_id,
   status).
-- **LanceDB** ([`infra/persistence/lancedb/tables/`](../src/everos/infra/persistence/lancedb/tables/))
-  holds the per-kind business rows, keyed `<owner_id>_<entry_id>` (so
-  cross-table joins use `(owner_id, entry_id)`); each table's `Vector(N)`
-  dimension matches the embedding model output.
-
-Episode and AtomicFact LanceDB tables carry a `deprecated_by: str | None`
-column. When an episode is superseded by a Reflection merge,
+- The **derived index backend** holds the per-kind business rows, keyed
+  `<owner_id>_<entry_id>` (so cross-table joins use `(owner_id, entry_id)`).
+  LanceDB is the default backend under `.index/lancedb/`; Milvus can be enabled
+  as the same rebuildable index backend and uses `.index/milvus/milvus.db` by
+  default for Milvus Lite.
+
+Episode and AtomicFact index rows carry a `deprecated_by: str | None` column.
+When an episode is superseded by a Reflection merge,
 `deprecated_by` is set to the merged episode's entry_id. Search filters
 automatically exclude rows where `deprecated_by IS NOT NULL`.
 
@@ -226,5 +231,5 @@ this primitive is **schema-agnostic** — field-level semantics
 - Code:
   - [`core/persistence/memory_root.py`](../src/everos/core/persistence/memory_root.py)
   - [`core/persistence/markdown/`](../src/everos/core/persistence/markdown/)
-  - [`infra/persistence/{markdown,sqlite,lancedb}/`](../src/everos/infra/persistence/)
-  - [`memory/cascade/`](../src/everos/memory/cascade/) (md → LanceDB sync)
+  - [`infra/persistence/{markdown,sqlite,lancedb,milvus,index}/`](../src/everos/infra/persistence/)
+  - [`memory/cascade/`](../src/everos/memory/cascade/) (md → derived index sync)
diff --git a/pyproject.toml b/pyproject.toml
index 0e359b6e6..86eac8594 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -8,7 +8,7 @@ requires-python = ">=3.12"
 authors = [
     {name = "EverMind AI"},
 ]
-keywords = ["memory", "ai-agent", "markdown", "lancedb", "rag", "everos"]
+keywords = ["memory", "ai-agent", "markdown", "lancedb", "milvus", "rag", "everos"]
 classifiers = [
     "Development Status :: 5 - Production/Stable",
     "Intended Audience :: Developers",
@@ -26,7 +26,7 @@ dependencies = [
     "pydantic-settings>=2.0.0",
 
     # Storage stack (md-first three-piece set)
-    "lancedb>=0.13.0",            # Vector + BM25 + scalar filter (Arrow-based)
+    "lancedb>=0.13.0",            # Default vector + BM25 + scalar filter index
     "aiosqlite>=0.20.0",          # Async SQLite driver (used by SA async engine)
     "sqlmodel>=0.0.22",           # ORM (Pydantic + SQLAlchemy 2.0 async)
     "alembic>=1.13.0",            # SQLite schema migrations
@@ -70,6 +70,9 @@ dependencies = [
 
 [project.optional-dependencies]
 multimodal = ["everalgo-parser[svg]>=0.2.1"]  # [svg] bundles cairosvg → SVG works by default
+milvus = [
+    "pymilvus[milvus-lite]>=3.0.0",
+]
 
 [project.urls]
 Homepage = "https://evermind.ai"
@@ -204,6 +207,7 @@ source_modules = ["everos.service", "everos.memory", "everos.entrypoints"]
 forbidden_modules = [
     "everos.infra.persistence.markdown.**",
     "everos.infra.persistence.lancedb.**",
+    "everos.infra.persistence.milvus.**",
     "everos.infra.persistence.sqlite.**",
 ]
 # `forbidden` contracts run a *transitive closure* — any path from a
@@ -222,12 +226,17 @@ forbidden_modules = [
 # in ``service`` or ``entrypoints``) remain blocked — that edge is not
 # in this whitelist.
 ignore_imports = [
+    "everos.infra.persistence.index -> everos.infra.persistence.lancedb",
+    "everos.infra.persistence.index -> everos.infra.persistence.milvus",
     "everos.infra.persistence.sqlite -> everos.infra.persistence.sqlite.tables",
     "everos.infra.persistence.sqlite -> everos.infra.persistence.sqlite.repos",
     "everos.infra.persistence.sqlite -> everos.infra.persistence.sqlite.sqlite_manager",
     "everos.infra.persistence.lancedb -> everos.infra.persistence.lancedb.tables",
     "everos.infra.persistence.lancedb -> everos.infra.persistence.lancedb.repos",
     "everos.infra.persistence.lancedb -> everos.infra.persistence.lancedb.lancedb_manager",
+    "everos.infra.persistence.milvus -> everos.infra.persistence.milvus.repos",
+    "everos.infra.persistence.milvus -> everos.infra.persistence.milvus.milvus_manager",
+    "everos.infra.persistence.milvus.repos -> everos.infra.persistence.milvus.repository",
     "everos.infra.persistence.markdown -> everos.infra.persistence.markdown.mds",
     "everos.infra.persistence.markdown -> everos.infra.persistence.markdown.writers",
     "everos.infra.persistence.markdown -> everos.infra.persistence.markdown.readers",
diff --git a/src/everos/README.md b/src/everos/README.md
index 8cc1cef96..2be6caf25 100644
--- a/src/everos/README.md
+++ b/src/everos/README.md
@@ -10,7 +10,7 @@ everos/
 ├── entrypoints/    Presentation: cli + api
 ├── service/        Application: use case orchestration
 ├── memory/         Domain: extract + search + cascade + prompt_slots + models
-├── infra/          Infrastructure: persistence/{markdown, sqlite, lancedb}
+├── infra/          Infrastructure: persistence/{markdown, sqlite, lancedb, milvus, index}
 ├── component/      Cross-cutting providers: llm / embedding / config / utils
 ├── core/           Runtime base: observability / lifespan / context
 └── config/         Data: Settings + default.toml + prompt_slots templates
diff --git a/src/everos/config/__init__.py b/src/everos/config/__init__.py
index c29590230..7ff8db9c9 100644
--- a/src/everos/config/__init__.py
+++ b/src/everos/config/__init__.py
@@ -2,7 +2,8 @@
 
 Public API:
     from everos.config import (
-        Settings, MemorySettings, SqliteSettings, LanceDBSettings,
+        Settings, MemorySettings, SqliteSettings, IndexSettings, LanceDBSettings,
+        MilvusSettings,
         LLMSettings, EmbeddingSettings, RerankSettings,
         BoundaryDetectionSettings,
         load_settings, resolve_root,
@@ -14,9 +15,11 @@
 
 from .settings import BoundaryDetectionSettings as BoundaryDetectionSettings
 from .settings import EmbeddingSettings as EmbeddingSettings
+from .settings import IndexSettings as IndexSettings
 from .settings import LanceDBSettings as LanceDBSettings
 from .settings import LLMSettings as LLMSettings
 from .settings import MemorySettings as MemorySettings
+from .settings import MilvusSettings as MilvusSettings
 from .settings import MultimodalSettings as MultimodalSettings
 from .settings import RerankSettings as RerankSettings
 from .settings import Settings as Settings
@@ -27,9 +30,11 @@
 __all__ = [
     "BoundaryDetectionSettings",
     "EmbeddingSettings",
+    "IndexSettings",
     "LLMSettings",
     "LanceDBSettings",
     "MemorySettings",
+    "MilvusSettings",
     "MultimodalSettings",
     "RerankSettings",
     "Settings",
diff --git a/src/everos/config/default.toml b/src/everos/config/default.toml
index 68fab81f1..9034ebcdf 100644
--- a/src/everos/config/default.toml
+++ b/src/everos/config/default.toml
@@ -50,6 +50,24 @@ cache_size_kb = 2048
 # Uncomment to override:
 # read_consistency_seconds = 5.0
 
+[index]
+# Rebuildable vector/BM25 index backend. Markdown remains the source of truth
+# and SQLite remains the system state store.
+# Override via EVEROS_INDEX__BACKEND.
+backend = "lancedb"
+
+[milvus]
+# Optional derived index backend. Leave uri empty to use Milvus Lite at
+# <root>/.index/milvus/milvus.db. Set uri/token/db_name for Milvus server or
+# Zilliz Cloud.
+# Override via EVEROS_MILVUS__URI, EVEROS_MILVUS__TOKEN, EVEROS_MILVUS__DB_NAME.
+uri = ""
+token = ""
+db_name = ""
+consistency_level = "Session"
+dimension = 1024
+collection_prefix = "everos"
+
 [llm]
 # Provider-agnostic OpenAI-protocol client config. Override via env:
 #   EVEROS_LLM__MODEL, EVEROS_LLM__API_KEY, EVEROS_LLM__BASE_URL
@@ -144,4 +162,3 @@ session_lock_timeout_seconds = 360.0
 # Override via EVEROS_CLUSTERING__THRESHOLD, EVEROS_CLUSTERING__TIME_WINDOW_DAYS.
 threshold = 0.65
 time_window_days = 7.0
-
diff --git a/src/everos/config/settings.py b/src/everos/config/settings.py
index 5ae2c8da8..5f9544491 100644
--- a/src/everos/config/settings.py
+++ b/src/everos/config/settings.py
@@ -331,6 +331,45 @@ class LanceDBSettings(BaseModel):
     index_cache_size_bytes: int = 16 * 1024 * 1024
 
 
+class IndexSettings(BaseModel):
+    """Derived index backend selection.
+
+    Markdown remains the source of truth and SQLite remains the system
+    state store. This setting only selects the rebuildable vector/BM25
+    index used by cascade and search.
+
+    Env binding:
+        EVEROS_INDEX__BACKEND
+    """
+
+    backend: Literal["lancedb", "milvus"] = "lancedb"
+
+
+class MilvusSettings(BaseModel):
+    """Milvus derived index connection and schema settings.
+
+    ``uri``:
+      ``None`` or empty string uses Milvus Lite under the memory root
+      (``<root>/.index/milvus/milvus.db``). Set to a Milvus server or
+      Zilliz Cloud endpoint to use an external service.
+
+    Env binding:
+        EVEROS_MILVUS__URI
+        EVEROS_MILVUS__TOKEN
+        EVEROS_MILVUS__DB_NAME
+        EVEROS_MILVUS__CONSISTENCY_LEVEL
+        EVEROS_MILVUS__DIMENSION
+        EVEROS_MILVUS__COLLECTION_PREFIX
+    """
+
+    uri: str | None = None
+    token: SecretStr | None = None
+    db_name: str | None = None
+    consistency_level: Literal["Strong", "Session", "Bounded", "Eventually"] = "Session"
+    dimension: int = Field(default=1024, ge=1)
+    collection_prefix: str = "everos"
+
+
 class KnowledgeSearchSettings(BaseModel):
     """``[knowledge.search]`` — retrieval tuning for the knowledge module."""
 
@@ -356,7 +395,9 @@ class Settings(BaseSettings):
     memory: MemorySettings = MemorySettings()
     api: ApiSettings = ApiSettings()
     sqlite: SqliteSettings = SqliteSettings()
+    index: IndexSettings = IndexSettings()
     lancedb: LanceDBSettings = LanceDBSettings()
+    milvus: MilvusSettings = MilvusSettings()
     llm: LLMSettings = LLMSettings()
     embedding: EmbeddingSettings = EmbeddingSettings()
     rerank: RerankSettings = RerankSettings()
diff --git a/src/everos/core/persistence/memory_root.py b/src/everos/core/persistence/memory_root.py
index 7a96bd2f9..efa507630 100644
--- a/src/everos/core/persistence/memory_root.py
+++ b/src/everos/core/persistence/memory_root.py
@@ -11,6 +11,7 @@
         .index/             derived indexes (rebuildable from markdown)
             sqlite/         system.db (+ WAL/SHM), ome.db, ome.aps.db
             lancedb/        LanceDB tables
+            milvus/         Milvus Lite database when Milvus is selected
         .tmp/               atomic-write staging directory
         .lock               single-process lock anchor (created on demand by
                             ``memory_root_lock``)
@@ -145,6 +146,16 @@ def lancedb_dir(self) -> Path:
         """``<root>/.index/lancedb/`` — LanceDB table root."""
         return self.index_dir / "lancedb"
 
+    @property
+    def milvus_dir(self) -> Path:
+        """``<root>/.index/milvus/`` — Milvus Lite derived index root."""
+        return self.index_dir / "milvus"
+
+    @property
+    def milvus_db(self) -> Path:
+        """``<root>/.index/milvus/milvus.db`` — default Milvus Lite URI."""
+        return self.milvus_dir / "milvus.db"
+
     @property
     def sqlite_dir(self) -> Path:
         """``<root>/.index/sqlite/`` — SQLite system DB root.
@@ -224,10 +235,12 @@ def ensure(self) -> None:
             <root>/.index/
             <root>/.index/sqlite/
             <root>/.index/lancedb/
+            <root>/.index/milvus/
             <root>/.tmp/
         """
         self.root.mkdir(parents=True, exist_ok=True)
         self.index_dir.mkdir(parents=True, exist_ok=True)
         self.sqlite_dir.mkdir(parents=True, exist_ok=True)
         self.lancedb_dir.mkdir(parents=True, exist_ok=True)
+        self.milvus_dir.mkdir(parents=True, exist_ok=True)
         self.tmp_dir.mkdir(parents=True, exist_ok=True)
diff --git a/src/everos/entrypoints/api/lifespans/lancedb.py b/src/everos/entrypoints/api/lifespans/lancedb.py
index b2a030bdc..0a6b00082 100644
--- a/src/everos/entrypoints/api/lifespans/lancedb.py
+++ b/src/everos/entrypoints/api/lifespans/lancedb.py
@@ -1,13 +1,11 @@
-"""LanceDB lifespan provider (HTTP API entrypoint).
+"""Derived index lifespan provider (HTTP API entrypoint).
 
 Startup:
-    Open the connection via ``get_connection`` (lazy, idempotent).
-    Importing :mod:`everos.infra.persistence.lancedb` also triggers the
-    side-effect import of ``tables`` so business schemas are loaded
-    (future: preflight registration).
+    Start the configured derived index backend (LanceDB by default,
+    Milvus when ``[index] backend = "milvus"``).
 
 Shutdown:
-    Close the connection (also clears the table cache).
+    Close the active backend connection.
 """
 
 from __future__ import annotations
@@ -18,38 +16,24 @@
 
 from everos.core.lifespan import LifespanProvider
 from everos.core.observability.logging import get_logger
-from everos.infra.persistence.lancedb import (
-    dispose_connection,
-    ensure_business_indexes,
-    get_connection,
-    verify_business_schemas,
-)
+from everos.infra.persistence.index import active_backend, shutdown, startup
 
 logger = get_logger(__name__)
 
 
 class LanceDBLifespanProvider(LifespanProvider):
-    """Manage the LanceDB connection + table cache for the app lifecycle.
+    """Manage the derived index backend for the app lifecycle.
 
-    Startup runs three steps:
-
-    1. ``get_connection`` — lazy-open the async connection.
-    2. ``verify_business_schemas`` — fail loud if an on-disk table's
-       columns drift from the current Pydantic schema. LanceDB has no
-       online migration; cascade is rebuildable from md so the recovery
-       is documented as ``rm -rf ~/.everos/.index/lancedb``.
-    3. ``ensure_business_indexes`` — idempotent FTS index creation.
+    The class name is kept for compatibility with existing app wiring.
     """
 
     def __init__(self, order: int = 11) -> None:
-        super().__init__(name="lancedb", order=order)
+        super().__init__(name="derived_index", order=order)
 
     async def startup(self, app: FastAPI) -> Any:
-        conn = await get_connection()
-        await verify_business_schemas()
-        await ensure_business_indexes()
-        logger.info("lancedb_ready", uri=conn.uri)
-        return conn
+        resource = await startup()
+        logger.info("derived_index_ready", backend=active_backend())
+        return resource
 
     async def shutdown(self, app: FastAPI) -> None:
-        await dispose_connection()
+        await shutdown()
diff --git a/src/everos/entrypoints/cli/commands/cascade.py b/src/everos/entrypoints/cli/commands/cascade.py
index 8d27e8aa0..328e25239 100644
--- a/src/everos/entrypoints/cli/commands/cascade.py
+++ b/src/everos/entrypoints/cli/commands/cascade.py
@@ -34,12 +34,8 @@
 from everos.component.utils.datetime import to_display_tz
 from everos.config import load_settings
 from everos.core.persistence import MemoryRoot
-from everos.infra.persistence.lancedb import (
-    dispose_connection,
-    ensure_business_indexes,
-    get_connection,
-    verify_business_schemas,
-)
+from everos.infra.persistence.index import shutdown as shutdown_index
+from everos.infra.persistence.index import startup as startup_index
 from everos.infra.persistence.sqlite import (
     dispose_engine,
     get_engine,
@@ -49,7 +45,7 @@
 
 app = typer.Typer(
     name="cascade",
-    help="Inspect and operate the md → LanceDB sync queue",
+    help="Inspect and operate the md → derived index sync queue",
     no_args_is_help=True,
 )
 
@@ -72,7 +68,7 @@ def _cascade_callback(
 
 @asynccontextmanager
 async def _runtime():  # type: ignore[no-untyped-def]
-    """Stand up sqlite + lancedb the same way the API lifespan would.
+    """Stand up sqlite + the derived index the same way the API lifespan would.
 
     The CLI piggybacks on the same singletons as the running daemon
     (lazy + process-wide), so if a server happens to be running on
@@ -81,13 +77,11 @@ async def _runtime():  # type: ignore[no-untyped-def]
     engine = get_engine()
     async with engine.begin() as conn:
         await conn.run_sync(SQLModel.metadata.create_all)
-    await get_connection()
-    await verify_business_schemas()
-    await ensure_business_indexes()
+    await startup_index()
     try:
         yield
     finally:
-        await dispose_connection()
+        await shutdown_index()
         await dispose_engine()
 
 
diff --git a/src/everos/infra/persistence/index/__init__.py b/src/everos/infra/persistence/index/__init__.py
new file mode 100644
index 000000000..0cc969eb6
--- /dev/null
+++ b/src/everos/infra/persistence/index/__init__.py
@@ -0,0 +1,327 @@
+"""Backend-neutral derived index persistence facade.
+
+Markdown is the source of truth and SQLite stores system state. This package
+selects the rebuildable vector/BM25 index backend used by cascade, search, and
+get. The default backend remains LanceDB; Milvus is opt-in through
+``[index] backend = "milvus"``.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import Any
+
+from lancedb.query import BooleanQuery, FullTextQuery, MatchQuery
+
+try:
+    from lancedb.query import Occur
+except ImportError:  # pragma: no cover
+    from lancedb._lancedb import Occur  # type: ignore[attr-defined,no-redef]
+
+from everos.config import load_settings
+from everos.infra.persistence import lancedb as _lancedb
+from everos.infra.persistence.lancedb import (
+    AgentCase,
+    AgentSkill,
+    AtomicFact,
+    Episode,
+    Foresight,
+    KnowledgeTopic,
+    ParentType,
+    UserProfile,
+)
+
+
+def active_backend() -> str:
+    """Return the configured derived index backend name."""
+    return load_settings().index.backend
+
+
+async def startup() -> Any:
+    """Initialise the configured derived index backend."""
+    if active_backend() == "milvus":
+        milvus = _milvus()
+        await milvus.get_client()
+        await milvus.ensure_business_indexes()
+        return "milvus"
+    conn = await _lancedb.get_connection()
+    await _lancedb.verify_business_schemas()
+    await _lancedb.ensure_business_indexes()
+    return conn
+
+
+async def shutdown() -> None:
+    """Dispose the configured derived index backend."""
+    if active_backend() == "milvus":
+        await _milvus().dispose_connection()
+    else:
+        await _lancedb.dispose_connection()
+
+
+async def ensure_business_indexes() -> None:
+    if active_backend() == "milvus":
+        await _milvus().ensure_business_indexes()
+    else:
+        await _lancedb.ensure_business_indexes()
+
+
+async def verify_business_schemas() -> None:
+    if active_backend() == "milvus":
+        await _milvus().verify_business_schemas()
+    else:
+        await _lancedb.verify_business_schemas()
+
+
+class _LanceIndexRepoAdapter:
+    """Add backend-neutral recall helpers to an existing LanceDB repo."""
+
+    def __init__(self, repo: Any, schema: type[Any]) -> None:
+        self._repo = repo
+        self.schema = schema
+
+    def __getattr__(self, name: str) -> Any:
+        return getattr(self._repo, name)
+
+    async def sparse_search(
+        self,
+        query_terms: Sequence[str],
+        where: Any,
+        *,
+        columns: Sequence[str] | None = None,
+        limit: int,
+    ) -> list[dict[str, Any]]:
+        fields = list(columns or self.schema.BM25_FIELDS)
+        if not query_terms or not fields:
+            return []
+        table = await _lancedb.get_table(self.schema.TABLE_NAME, self.schema)
+        expr = _lance_expr(where)
+        best: dict[str, dict[str, Any]] = {}
+        for field in fields:
+            query = _build_or_query(query_terms, field)
+            rows = (
+                await table.query()
+                .nearest_to_text(query)
+                .where(expr)
+                .limit(limit)
+                .to_list()
+            )
+            for row in rows:
+                rid = row.get("id")
+                if not isinstance(rid, str):
+                    continue
+                score = float(row.get("_score", 0.0))
+                prior = best.get(rid)
+                if prior is None or score > float(prior.get("_score", 0.0)):
+                    shaped = dict(row)
+                    shaped["_score"] = score
+                    best[rid] = shaped
+        return sorted(
+            best.values(), key=lambda row: float(row.get("_score", 0.0)), reverse=True
+        )[:limit]
+
+    async def dense_search(
+        self,
+        vector: Sequence[float],
+        where: Any,
+        *,
+        limit: int,
+    ) -> list[dict[str, Any]]:
+        if not vector:
+            return []
+        table = await _lancedb.get_table(self.schema.TABLE_NAME, self.schema)
+        return (
+            await table.query()
+            .nearest_to(list(vector))
+            .distance_type("cosine")
+            .where(_lance_expr(where))
+            .limit(limit)
+            .to_list()
+        )
+
+
+class _IndexRepoRouter:
+    """Route repo calls to the configured derived index backend."""
+
+    def __init__(
+        self, lance_repo: Any, milvus_repo_name: str, schema: type[Any]
+    ) -> None:
+        self._lance = _LanceIndexRepoAdapter(lance_repo, schema)
+        self._milvus_repo_name = milvus_repo_name
+        self.schema = schema
+
+    @property
+    def table_name(self) -> str:
+        return self.schema.TABLE_NAME
+
+    def _repo(self) -> Any:
+        if active_backend() == "milvus":
+            return getattr(_milvus(), self._milvus_repo_name)
+        return self._lance
+
+    def __getattr__(self, name: str) -> Any:
+        return getattr(self._repo(), name)
+
+    async def add(self, records: Sequence[Any]) -> None:
+        await self._repo().add(records)
+
+    async def upsert(self, records: Sequence[Any], *, by: str = "id") -> None:
+        await self._repo().upsert(records, by=by)
+
+    async def count(self) -> int:
+        return await self._repo().count()
+
+    async def get_by_id(self, id_value: str, *, id_field: str = "id") -> Any:
+        return await self._repo().get_by_id(id_value, id_field=id_field)
+
+    async def find_where(self, where: Any, *, limit: int = 100) -> list[Any]:
+        return await self._repo().find_where(_where_for_backend(where), limit=limit)
+
+    async def find_one_where(self, where: Any) -> Any:
+        return await self._repo().find_one_where(_where_for_backend(where))
+
+    async def find_where_paginated(
+        self,
+        where: Any,
+        *,
+        sort_by: str,
+        descending: bool = True,
+        page: int = 1,
+        page_size: int = 20,
+        max_fetch: int = 20_000,
+    ) -> tuple[list[Any], int]:
+        return await self._repo().find_where_paginated(
+            _where_for_backend(where),
+            sort_by=sort_by,
+            descending=descending,
+            page=page,
+            page_size=page_size,
+            max_fetch=max_fetch,
+        )
+
+    async def search(
+        self,
+        *,
+        vector: Sequence[float] | None = None,
+        where: Any = None,
+        limit: int = 10,
+    ) -> list[dict[str, Any]]:
+        return await self._repo().search(
+            vector=vector, where=_where_for_backend(where), limit=limit
+        )
+
+    async def update(self, updates: dict[str, Any], *, where: Any) -> None:
+        await self._repo().update(updates, where=_where_for_backend(where))
+
+    async def delete(self, predicate: Any) -> None:
+        await self._repo().delete(_where_for_backend(predicate))
+
+    async def delete_by_md_path(self, md_path: str) -> int:
+        return await self._repo().delete_by_md_path(md_path)
+
+    async def optimize(self, **kwargs: Any) -> None:
+        await self._repo().optimize(**kwargs)
+
+    async def rebuild_indexes(self) -> None:
+        await self._repo().rebuild_indexes()
+
+    async def sparse_search(
+        self,
+        query_terms: Sequence[str],
+        where: Any,
+        *,
+        columns: Sequence[str] | None = None,
+        limit: int,
+    ) -> list[dict[str, Any]]:
+        return await self._repo().sparse_search(
+            query_terms,
+            _where_for_backend(where),
+            columns=columns,
+            limit=limit,
+        )
+
+    async def dense_search(
+        self, vector: Sequence[float], where: Any, *, limit: int
+    ) -> list[dict[str, Any]]:
+        return await self._repo().dense_search(
+            vector, _where_for_backend(where), limit=limit
+        )
+
+
+def _where_for_backend(where: Any) -> Any:
+    if active_backend() == "milvus":
+        return getattr(where, "milvus", where)
+    return _lance_expr(where)
+
+
+def _lance_expr(where: Any) -> str:
+    return getattr(where, "lancedb", where)
+
+
+def _build_or_query(tokens: Sequence[str], column: str) -> FullTextQuery:
+    clean = [token for token in tokens if token]
+    if not clean:
+        return MatchQuery("", column=column)
+    if len(clean) == 1:
+        return MatchQuery(clean[0], column=column)
+    return BooleanQuery(
+        [(Occur.SHOULD, MatchQuery(token, column=column)) for token in clean]
+    )
+
+
+def _milvus() -> Any:
+    from everos.infra.persistence import milvus
+
+    return milvus
+
+
+episode_repo = _IndexRepoRouter(_lancedb.episode_repo, "episode_repo", Episode)
+atomic_fact_repo = _IndexRepoRouter(
+    _lancedb.atomic_fact_repo, "atomic_fact_repo", AtomicFact
+)
+foresight_repo = _IndexRepoRouter(_lancedb.foresight_repo, "foresight_repo", Foresight)
+agent_case_repo = _IndexRepoRouter(
+    _lancedb.agent_case_repo, "agent_case_repo", AgentCase
+)
+agent_skill_repo = _IndexRepoRouter(
+    _lancedb.agent_skill_repo, "agent_skill_repo", AgentSkill
+)
+user_profile_repo = _IndexRepoRouter(
+    _lancedb.user_profile_repo, "user_profile_repo", UserProfile
+)
+knowledge_topic_repo = _IndexRepoRouter(
+    _lancedb.knowledge_topic_repo, "knowledge_topic_repo", KnowledgeTopic
+)
+
+ALL_REPOS = (
+    episode_repo,
+    atomic_fact_repo,
+    foresight_repo,
+    agent_case_repo,
+    agent_skill_repo,
+    user_profile_repo,
+    knowledge_topic_repo,
+)
+
+__all__ = [
+    "ALL_REPOS",
+    "AgentCase",
+    "AgentSkill",
+    "AtomicFact",
+    "Episode",
+    "Foresight",
+    "KnowledgeTopic",
+    "ParentType",
+    "UserProfile",
+    "active_backend",
+    "agent_case_repo",
+    "agent_skill_repo",
+    "atomic_fact_repo",
+    "ensure_business_indexes",
+    "episode_repo",
+    "foresight_repo",
+    "knowledge_topic_repo",
+    "shutdown",
+    "startup",
+    "user_profile_repo",
+    "verify_business_schemas",
+]
diff --git a/src/everos/infra/persistence/milvus/__init__.py b/src/everos/infra/persistence/milvus/__init__.py
new file mode 100644
index 000000000..4396360c8
--- /dev/null
+++ b/src/everos/infra/persistence/milvus/__init__.py
@@ -0,0 +1,49 @@
+"""Milvus derived index backend.
+
+This package mirrors the LanceDB business index surface but stores rows in
+Milvus collections. It is selected through ``Settings.index.backend`` and is
+normally reached through :mod:`everos.infra.persistence.index`.
+"""
+
+from __future__ import annotations
+
+from .milvus_manager import MilvusSchemaMismatchError as MilvusSchemaMismatchError
+from .milvus_manager import dispose_connection as dispose_connection
+from .milvus_manager import get_client as get_client
+from .repos import ALL_REPOS as ALL_REPOS
+from .repos import agent_case_repo as agent_case_repo
+from .repos import agent_skill_repo as agent_skill_repo
+from .repos import atomic_fact_repo as atomic_fact_repo
+from .repos import episode_repo as episode_repo
+from .repos import foresight_repo as foresight_repo
+from .repos import knowledge_topic_repo as knowledge_topic_repo
+from .repos import user_profile_repo as user_profile_repo
+
+
+async def ensure_business_indexes() -> None:
+    """Create or verify every EverOS Milvus collection."""
+    for repo in ALL_REPOS:
+        await repo.ensure_collection()
+
+
+async def verify_business_schemas() -> None:
+    """Fail loud if existing Milvus collections drift from EverOS schemas."""
+    for repo in ALL_REPOS:
+        await repo.verify_collection()
+
+
+__all__ = [
+    "ALL_REPOS",
+    "MilvusSchemaMismatchError",
+    "agent_case_repo",
+    "agent_skill_repo",
+    "atomic_fact_repo",
+    "dispose_connection",
+    "ensure_business_indexes",
+    "episode_repo",
+    "foresight_repo",
+    "get_client",
+    "knowledge_topic_repo",
+    "user_profile_repo",
+    "verify_business_schemas",
+]
diff --git a/src/everos/infra/persistence/milvus/milvus_manager.py b/src/everos/infra/persistence/milvus/milvus_manager.py
new file mode 100644
index 000000000..602e4a314
--- /dev/null
+++ b/src/everos/infra/persistence/milvus/milvus_manager.py
@@ -0,0 +1,87 @@
+"""Milvus connection and collection management for the derived index."""
+
+from __future__ import annotations
+
+import re
+
+from pymilvus import MilvusClient
+
+from everos.config import MilvusSettings, load_settings
+from everos.core.observability.logging import get_logger
+from everos.core.persistence import MemoryRoot
+
+logger = get_logger(__name__)
+
+_client: MilvusClient | None = None
+
+
+class MilvusSchemaMismatchError(RuntimeError):
+    """Raised when an existing Milvus collection does not match EverOS."""
+
+
+def collection_name(table_name: str, settings: MilvusSettings | None = None) -> str:
+    """Return the configured Milvus collection name for an EverOS table."""
+    cfg = settings or load_settings().milvus
+    prefix = _sanitize_name_part(cfg.collection_prefix)
+    base = _sanitize_name_part(table_name)
+    name = f"{prefix}_{base}" if prefix else base
+    if not re.match(r"^[A-Za-z_]", name):
+        name = f"_{name}"
+    return name
+
+
+async def get_client() -> MilvusClient:
+    """Return the process-wide MilvusClient, creating it lazily."""
+    global _client
+    if _client is None:
+        settings = load_settings().milvus
+        uri = _resolve_uri(settings)
+        token = _secret(settings.token)
+        db_name = settings.db_name or ""
+        _client = MilvusClient(uri=uri, token=token, db_name=db_name)
+        logger.info(
+            "milvus_connection_opened",
+            uri=uri,
+            db_name=db_name or None,
+            consistency_level=settings.consistency_level,
+        )
+    return _client
+
+
+async def dispose_connection() -> None:
+    """Close the process-wide Milvus client."""
+    global _client
+    if _client is not None:
+        _client.close()
+        _client = None
+        logger.info("milvus_connection_closed")
+
+
+def _resolve_uri(settings: MilvusSettings) -> str:
+    if settings.uri:
+        return settings.uri
+    memory_root = MemoryRoot.default()
+    memory_root.milvus_dir.mkdir(parents=True, exist_ok=True)
+    return str(memory_root.milvus_db)
+
+
+def _secret(value: object | None) -> str:
+    if value is None:
+        return ""
+    getter = getattr(value, "get_secret_value", None)
+    if callable(getter):
+        return getter() or ""
+    return str(value)
+
+
+def _sanitize_name_part(value: str) -> str:
+    clean = re.sub(r"\W+", "_", value.strip())
+    return clean.strip("_")
+
+
+__all__ = [
+    "MilvusSchemaMismatchError",
+    "collection_name",
+    "dispose_connection",
+    "get_client",
+]
diff --git a/src/everos/infra/persistence/milvus/repos.py b/src/everos/infra/persistence/milvus/repos.py
new file mode 100644
index 000000000..826bc9267
--- /dev/null
+++ b/src/everos/infra/persistence/milvus/repos.py
@@ -0,0 +1,119 @@
+"""Milvus repo singletons for EverOS derived index tables."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from everos.infra.persistence.lancedb import (
+    AgentCase,
+    AgentSkill,
+    AtomicFact,
+    Episode,
+    Foresight,
+    KnowledgeTopic,
+    UserProfile,
+)
+
+from .repository import MilvusRepoBase
+
+
+class _EpisodeRepo(MilvusRepoBase[Episode]):
+    schema = Episode
+
+
+class _AtomicFactRepo(MilvusRepoBase[AtomicFact]):
+    schema = AtomicFact
+
+
+class _ForesightRepo(MilvusRepoBase[Foresight]):
+    schema = Foresight
+
+
+class _AgentCaseRepo(MilvusRepoBase[AgentCase]):
+    schema = AgentCase
+
+
+class _AgentSkillRepo(MilvusRepoBase[AgentSkill]):
+    schema = AgentSkill
+
+    async def count_in_cluster(self, *, owner_id: str, cluster_id: str) -> int:
+        rows = await self.find_where(
+            f"owner_id = '{_q(owner_id)}' AND cluster_id = '{_q(cluster_id)}'",
+            limit=10_000,
+        )
+        return len(rows)
+
+    async def find_in_cluster(
+        self, *, owner_id: str, cluster_id: str, limit: int
+    ) -> list[AgentSkill]:
+        return await self.find_where(
+            f"owner_id = '{_q(owner_id)}' AND cluster_id = '{_q(cluster_id)}'",
+            limit=limit,
+        )
+
+    async def find_topk_relevant_in_cluster(
+        self,
+        *,
+        owner_id: str,
+        cluster_id: str,
+        query_vector: Sequence[float],
+        top_k: int,
+    ) -> list[AgentSkill]:
+        if not query_vector:
+            raise ValueError(
+                "query_vector must be non-empty; "
+                "call find_in_cluster for the scalar fallback"
+            )
+        rows = await self.dense_search(
+            query_vector,
+            f"owner_id = '{_q(owner_id)}' AND cluster_id = '{_q(cluster_id)}'",
+            limit=top_k,
+        )
+        out: list[AgentSkill] = []
+        for row in rows:
+            rid = row.get("id")
+            if isinstance(rid, str) and (item := await self.get_by_id(rid)) is not None:
+                out.append(item)
+        return out
+
+
+class _UserProfileRepo(MilvusRepoBase[UserProfile]):
+    schema = UserProfile
+
+
+class _KnowledgeTopicRepo(MilvusRepoBase[KnowledgeTopic]):
+    schema = KnowledgeTopic
+
+
+def _q(value: str) -> str:
+    return value.replace("\\", "\\\\").replace("'", "\\'")
+
+
+episode_repo = _EpisodeRepo()
+atomic_fact_repo = _AtomicFactRepo()
+foresight_repo = _ForesightRepo()
+agent_case_repo = _AgentCaseRepo()
+agent_skill_repo = _AgentSkillRepo()
+user_profile_repo = _UserProfileRepo()
+knowledge_topic_repo = _KnowledgeTopicRepo()
+
+ALL_REPOS = (
+    episode_repo,
+    atomic_fact_repo,
+    foresight_repo,
+    agent_case_repo,
+    agent_skill_repo,
+    user_profile_repo,
+    knowledge_topic_repo,
+)
+
+__all__ = [
+    "ALL_REPOS",
+    "agent_case_repo",
+    "agent_skill_repo",
+    "atomic_fact_repo",
+    "episode_repo",
+    "foresight_repo",
+    "knowledge_topic_repo",
+    "user_profile_repo",
+]
diff --git a/src/everos/infra/persistence/milvus/repository.py b/src/everos/infra/persistence/milvus/repository.py
new file mode 100644
index 000000000..de9143b3d
--- /dev/null
+++ b/src/everos/infra/persistence/milvus/repository.py
@@ -0,0 +1,758 @@
+"""Milvus repository implementation for EverOS derived index tables."""
+
+from __future__ import annotations
+
+import asyncio
+import datetime as dt
+import re
+from collections.abc import Sequence
+from importlib.metadata import PackageNotFoundError, version
+from typing import Any, ClassVar, get_args, get_origin
+
+from pymilvus import DataType, Function, FunctionType, MilvusClient
+
+from everos.component.utils.datetime import (
+    ensure_utc,
+    from_iso_format,
+    from_timestamp,
+    to_timestamp_ms,
+)
+from everos.config import load_settings
+from everos.core.observability.logging import get_logger
+from everos.core.persistence import BaseLanceTable
+
+from .milvus_manager import (
+    MilvusSchemaMismatchError,
+    collection_name,
+    get_client,
+)
+
+logger = get_logger(__name__)
+
+_DUMMY_VECTOR_FIELD = "_everos_dummy_vector"
+_SPARSE_SUFFIX = "__sparse"
+_MAX_VARCHAR_LENGTH = 65_535
+_ID_MAX_LENGTH = 512
+_ARRAY_MAX_CAPACITY = 256
+
+
+def _q(value: str) -> str:
+    """Escape a string for a Milvus single-quoted expression literal."""
+    return value.replace("\\", "\\\\").replace("'", "\\'")
+
+
+class MilvusRepoBase[T: BaseLanceTable]:
+    """Generic Milvus repository for one EverOS derived index collection."""
+
+    schema: type[T]
+    _write_locks: ClassVar[dict[str, asyncio.Lock]] = {}
+
+    @property
+    def table_name(self) -> str:
+        return self.schema.TABLE_NAME
+
+    @property
+    def collection_name(self) -> str:
+        return collection_name(self.table_name)
+
+    @classmethod
+    def _write_lock(cls, name: str) -> asyncio.Lock:
+        return cls._write_locks.setdefault(name, asyncio.Lock())
+
+    @classmethod
+    def _reset_locks_for_tests(cls) -> None:
+        cls._write_locks.clear()
+
+    async def ensure_collection(self) -> None:
+        client = await get_client()
+        if await _run(client.has_collection, self.collection_name):
+            await self.verify_collection()
+            return
+
+        schema = self._build_collection_schema()
+        index_params = client.prepare_index_params()
+        if "vector" in self.schema.model_fields:
+            index_params.add_index(
+                field_name="vector",
+                index_type="AUTOINDEX",
+                metric_type="COSINE",
+            )
+        else:
+            index_params.add_index(
+                field_name=_DUMMY_VECTOR_FIELD,
+                index_type="AUTOINDEX",
+                metric_type="COSINE",
+            )
+        for field in self.schema.BM25_FIELDS:
+            index_params.add_index(
+                field_name=_sparse_field(field),
+                index_type="AUTOINDEX",
+                metric_type="BM25",
+            )
+
+        settings = load_settings().milvus
+        await _run(
+            client.create_collection,
+            collection_name=self.collection_name,
+            schema=schema,
+            index_params=index_params,
+            consistency_level=settings.consistency_level,
+        )
+        logger.info(
+            "milvus_collection_created",
+            table=self.table_name,
+            collection=self.collection_name,
+        )
+
+    async def verify_collection(self) -> None:
+        client = await get_client()
+        description = await _run(client.describe_collection, self.collection_name)
+        actual = {field["name"] for field in description.get("fields", [])}
+        expected = set(self._stored_field_names())
+        missing = expected - actual
+        if missing:
+            raise MilvusSchemaMismatchError(
+                f"Milvus collection {self.collection_name!r} schema drift: "
+                f"missing={sorted(missing)}. The index is rebuildable from md; "
+                "drop the collection and restart to rebuild it."
+            )
+
+    # ── Create / update ─────────────────────────────────────────────
+
+    async def add(self, records: Sequence[T]) -> None:
+        if not records:
+            return
+        await self.ensure_collection()
+        client = await get_client()
+        payload = [self._to_milvus_record(record) for record in records]
+        async with self._write_lock(self.collection_name):
+            await _run(client.insert, self.collection_name, payload)
+
+    async def upsert(self, records: Sequence[T], *, by: str = "id") -> None:
+        if by != "id":
+            raise ValueError("MilvusRepoBase only supports upsert by id")
+        if not records:
+            return
+        await self.ensure_collection()
+        client = await get_client()
+        payload = [self._to_milvus_record(record) for record in records]
+        async with self._write_lock(self.collection_name):
+            await _run(client.upsert, self.collection_name, payload)
+
+    async def update(self, updates: dict[str, Any], *, where: Any) -> None:
+        await self.ensure_collection()
+        rows = await self._query_raw(where, limit=10_000, include_vectors=True)
+        if not rows:
+            return
+        patched = []
+        for row in rows:
+            merged = dict(row)
+            for key, value in updates.items():
+                self._write_field_value(merged, key, value)
+            patched.append(merged)
+        client = await get_client()
+        async with self._write_lock(self.collection_name):
+            await _run(client.upsert, self.collection_name, patched)
+
+    # ── Maintenance ────────────────────────────────────────────────
+
+    async def optimize(self, *, cleanup_older_than: dt.timedelta | None = None) -> None:
+        """Milvus indexes are managed by the service; no per-write compaction."""
+
+    async def rebuild_indexes(self) -> None:
+        """Milvus AUTOINDEX maintenance is service-managed."""
+
+    # ── Reads ──────────────────────────────────────────────────────
+
+    async def count(self) -> int:
+        await self.ensure_collection()
+        client = await get_client()
+        rows = await _run(
+            client.query,
+            self.collection_name,
+            filter="",
+            output_fields=["count(*)"],
+        )
+        if not rows:
+            return 0
+        return int(rows[0].get("count(*)", 0))
+
+    async def get_by_id(self, id_value: str, *, id_field: str = "id") -> T | None:
+        if id_field != "id":
+            rows = await self.find_where(f"{id_field} = '{_q(id_value)}'", limit=1)
+            return rows[0] if rows else None
+        await self.ensure_collection()
+        client = await get_client()
+        rows = await _run(
+            client.get,
+            self.collection_name,
+            ids=[id_value],
+            output_fields=self._output_fields(include_vectors=True),
+        )
+        if not rows:
+            return None
+        return self._model_from_milvus(rows[0])
+
+    async def find_where(self, where: Any, *, limit: int = 100) -> list[T]:
+        rows = await self._query_raw(where, limit=limit, include_vectors=True)
+        return [self._model_from_milvus(row) for row in rows]
+
+    async def find_one_where(self, where: Any) -> T | None:
+        rows = await self.find_where(where, limit=1)
+        return rows[0] if rows else None
+
+    async def find_where_paginated(
+        self,
+        where: Any,
+        *,
+        sort_by: str,
+        descending: bool = True,
+        page: int = 1,
+        page_size: int = 20,
+        max_fetch: int = 20_000,
+    ) -> tuple[list[T], int]:
+        raw = await self._query_raw(where, limit=max_fetch, include_vectors=True)
+        total = len(raw)
+        if total >= max_fetch:
+            logger.warning(
+                "milvus_find_where_paginated_truncated",
+                table=self.table_name,
+                total=total,
+                max_fetch=max_fetch,
+            )
+        rows = [self._model_from_milvus(row) for row in raw]
+        rows.sort(
+            key=lambda row: _sort_value(getattr(row, sort_by, None)),
+            reverse=descending,
+        )
+        offset = (page - 1) * page_size
+        return rows[offset : offset + page_size], total
+
+    async def find_by_owner(self, owner_id: str, *, limit: int = 100) -> list[T]:
+        return await self.find_where(f"owner_id = '{_q(owner_id)}'", limit=limit)
+
+    async def find_by_md_path(self, md_path: str) -> T | None:
+        return await self.find_one_where(f"md_path = '{_q(md_path)}'")
+
+    async def search(
+        self,
+        *,
+        vector: Sequence[float] | None = None,
+        where: Any = None,
+        limit: int = 10,
+    ) -> list[dict[str, Any]]:
+        if vector is None:
+            return await self._query_candidate_rows(where, limit=limit)
+        return await self.dense_search(vector, where, limit=limit)
+
+    async def sparse_search(
+        self,
+        query_terms: Sequence[str],
+        where: Any,
+        *,
+        columns: Sequence[str] | None = None,
+        limit: int,
+    ) -> list[dict[str, Any]]:
+        if not query_terms:
+            return []
+        await self.ensure_collection()
+        fields = list(columns or self.schema.BM25_FIELDS)
+        if not fields:
+            return []
+        client = await get_client()
+        expr = self._expr(where)
+        query = " ".join(term for term in query_terms if term)
+        best: dict[str, dict[str, Any]] = {}
+        for field in fields:
+            results = await _run(
+                client.search,
+                self.collection_name,
+                data=[query],
+                anns_field=_sparse_field(field),
+                filter=expr,
+                limit=limit,
+                output_fields=self._output_fields(include_vectors=False),
+                search_params={"metric_type": "BM25"},
+            )
+            for row in _first_result_set(results):
+                shaped = self._candidate_row_from_search(row)
+                score = _bm25_score_from_distance(row.get("distance"))
+                shaped["_score"] = score
+                rid = shaped.get("id")
+                if not isinstance(rid, str):
+                    continue
+                prior = best.get(rid)
+                if prior is None or score > float(prior.get("_score", 0.0)):
+                    best[rid] = shaped
+        return sorted(
+            best.values(), key=lambda item: float(item.get("_score", 0.0)), reverse=True
+        )[:limit]
+
+    async def dense_search(
+        self,
+        vector: Sequence[float],
+        where: Any,
+        *,
+        limit: int,
+    ) -> list[dict[str, Any]]:
+        if not vector or "vector" not in self.schema.model_fields:
+            return []
+        await self.ensure_collection()
+        client = await get_client()
+        results = await _run(
+            client.search,
+            self.collection_name,
+            data=[list(vector)],
+            anns_field="vector",
+            filter=self._expr(where),
+            limit=limit,
+            output_fields=self._output_fields(include_vectors=False),
+            search_params={"metric_type": "COSINE"},
+        )
+        return [
+            self._candidate_row_from_search(row, normalize_cosine=True)
+            for row in _first_result_set(results)
+        ]
+
+    # ── Deletes ────────────────────────────────────────────────────
+
+    async def delete(self, predicate: Any) -> None:
+        await self.ensure_collection()
+        client = await get_client()
+        async with self._write_lock(self.collection_name):
+            await _run(
+                client.delete,
+                self.collection_name,
+                filter=self._expr(predicate),
+            )
+
+    async def delete_by_md_path(self, md_path: str) -> int:
+        await self.ensure_collection()
+        before = await self.find_where(f"md_path = '{_q(md_path)}'", limit=10_000)
+        if not before:
+            return 0
+        await self.delete(f"md_path = '{_q(md_path)}'")
+        return len(before)
+
+    # ── Daily-log helpers ──────────────────────────────────────────
+
+    async def find_by_owner_entry(
+        self,
+        owner_id: str,
+        entry_id: str,
+        *,
+        app_id: str = "default",
+        project_id: str = "default",
+    ) -> T | None:
+        return await self.find_one_where(
+            f"owner_id = '{_q(owner_id)}' AND entry_id = '{_q(entry_id)}' "
+            f"AND app_id = '{_q(app_id)}' AND project_id = '{_q(project_id)}'"
+        )
+
+    async def find_by_owner_entries(
+        self,
+        owner_id: str,
+        entry_ids: Sequence[str],
+        *,
+        app_id: str = "default",
+        project_id: str = "default",
+    ) -> list[T]:
+        if not entry_ids:
+            return []
+        quoted = ", ".join(f"'{_q(entry_id)}'" for entry_id in entry_ids)
+        return await self.find_where(
+            f"owner_id = '{_q(owner_id)}' AND entry_id IN ({quoted}) "
+            f"AND app_id = '{_q(app_id)}' AND project_id = '{_q(project_id)}'",
+            limit=len(entry_ids),
+        )
+
+    async def find_by_session(
+        self, owner_id: str, session_id: str, *, limit: int = 100
+    ) -> list[T]:
+        return await self.find_where(
+            f"owner_id = '{_q(owner_id)}' AND session_id = '{_q(session_id)}'",
+            limit=limit,
+        )
+
+    async def find_by_parent(
+        self, parent_type: str, parent_id: str, *, limit: int = 100
+    ) -> list[T]:
+        return await self.find_where(
+            f"parent_type = '{_q(parent_type)}' AND parent_id = '{_q(parent_id)}'",
+            limit=limit,
+        )
+
+    # ── Field conversion ───────────────────────────────────────────
+
+    def _build_collection_schema(self):  # type: ignore[no-untyped-def]
+        settings = load_settings().milvus
+        schema = MilvusClient.create_schema(auto_id=False, enable_dynamic_field=False)
+        has_vector = False
+        for name, field in self.schema.model_fields.items():
+            if name == "subject_vector":
+                continue
+            annotation = field.annotation
+            if name == "id":
+                schema.add_field(
+                    field_name=name,
+                    datatype=DataType.VARCHAR,
+                    is_primary=True,
+                    max_length=_ID_MAX_LENGTH,
+                )
+            elif name == "vector":
+                has_vector = True
+                schema.add_field(
+                    field_name=name,
+                    datatype=DataType.FLOAT_VECTOR,
+                    dim=settings.dimension,
+                )
+            elif _is_datetime(annotation):
+                schema.add_field(
+                    field_name=_datetime_storage_field(name),
+                    datatype=DataType.INT64,
+                    nullable=_is_optional(annotation),
+                )
+            elif _is_list(annotation):
+                schema.add_field(
+                    field_name=name,
+                    datatype=DataType.ARRAY,
+                    element_type=DataType.VARCHAR,
+                    max_capacity=_ARRAY_MAX_CAPACITY,
+                    max_length=_ID_MAX_LENGTH,
+                    nullable=_is_optional(annotation),
+                )
+            elif _is_float(annotation):
+                schema.add_field(
+                    field_name=name,
+                    datatype=DataType.DOUBLE,
+                    nullable=_is_optional(annotation),
+                )
+            elif _is_int(annotation):
+                schema.add_field(
+                    field_name=name,
+                    datatype=DataType.INT64,
+                    nullable=_is_optional(annotation),
+                )
+            else:
+                kwargs: dict[str, Any] = {
+                    "field_name": name,
+                    "datatype": DataType.VARCHAR,
+                    "max_length": _MAX_VARCHAR_LENGTH,
+                }
+                if name in self.schema.BM25_FIELDS:
+                    kwargs["enable_analyzer"] = True
+                elif _is_optional(annotation):
+                    kwargs["nullable"] = True
+                schema.add_field(**kwargs)
+
+        if not has_vector:
+            schema.add_field(
+                field_name=_DUMMY_VECTOR_FIELD,
+                datatype=DataType.FLOAT_VECTOR,
+                dim=1,
+            )
+        for field in self.schema.BM25_FIELDS:
+            schema.add_field(
+                field_name=_sparse_field(field),
+                datatype=DataType.SPARSE_FLOAT_VECTOR,
+            )
+            schema.add_function(
+                Function(
+                    name=f"{field}_bm25",
+                    function_type=FunctionType.BM25,
+                    input_field_names=[field],
+                    output_field_names=[_sparse_field(field)],
+                )
+            )
+        return schema
+
+    def _stored_field_names(self) -> list[str]:
+        names: list[str] = []
+        has_vector = False
+        for name, field in self.schema.model_fields.items():
+            if name == "subject_vector":
+                continue
+            if name == "vector":
+                has_vector = True
+                names.append(name)
+            elif _is_datetime(field.annotation):
+                names.append(_datetime_storage_field(name))
+            else:
+                names.append(name)
+        if not has_vector:
+            names.append(_DUMMY_VECTOR_FIELD)
+        names.extend(_sparse_field(field) for field in self.schema.BM25_FIELDS)
+        return names
+
+    def _output_fields(self, *, include_vectors: bool) -> list[str]:
+        fields: list[str] = []
+        for name in self._stored_field_names():
+            if name.endswith(_SPARSE_SUFFIX) or name == _DUMMY_VECTOR_FIELD:
+                continue
+            if not include_vectors and name == "vector":
+                continue
+            fields.append(name)
+        return fields
+
+    def _to_milvus_record(self, record: T) -> dict[str, Any]:
+        raw = record.model_dump(mode="python")
+        out: dict[str, Any] = {}
+        for name, field in self.schema.model_fields.items():
+            if name == "subject_vector":
+                continue
+            value = raw.get(name)
+            if name in self.schema.BM25_FIELDS and value is None:
+                value = ""
+            if _is_datetime(field.annotation):
+                storage_name = _datetime_storage_field(name)
+                out[storage_name] = (
+                    _datetime_to_ms(value) if value is not None else None
+                )
+            elif name == "vector":
+                out[name] = list(value or [])
+            elif _is_list(field.annotation):
+                out[name] = [str(item) for item in (value or [])]
+            else:
+                out[name] = value
+        if "vector" not in self.schema.model_fields:
+            out[_DUMMY_VECTOR_FIELD] = [0.0]
+        return out
+
+    def _model_from_milvus(self, row: dict[str, Any]) -> T:
+        shaped = self._restore_row(row, include_distance=False)
+        return self.schema.model_validate(shaped)
+
+    def _candidate_row_from_search(
+        self, row: dict[str, Any], *, normalize_cosine: bool = False
+    ) -> dict[str, Any]:
+        shaped = self._restore_row(row.get("entity", {}), include_distance=False)
+        raw_distance = row.get("distance")
+        shaped["_distance"] = (
+            _cosine_distance_from_milvus(raw_distance)
+            if normalize_cosine
+            else raw_distance
+        )
+        return shaped
+
+    def _restore_row(
+        self, row: dict[str, Any], *, include_distance: bool
+    ) -> dict[str, Any]:
+        out: dict[str, Any] = {}
+        for name, field in self.schema.model_fields.items():
+            if name == "subject_vector":
+                continue
+            if _is_datetime(field.annotation):
+                storage_name = _datetime_storage_field(name)
+                value = row.get(storage_name)
+                if value is None:
+                    out[name] = None
+                else:
+                    out[name] = from_timestamp(int(value))
+            elif name in row:
+                value = row[name]
+                if (
+                    name in self.schema.BM25_FIELDS
+                    and value == ""
+                    and _is_optional(field.annotation)
+                ):
+                    out[name] = None
+                else:
+                    out[name] = value
+        if include_distance and "distance" in row:
+            out["_distance"] = row["distance"]
+        return out
+
+    def _write_field_value(
+        self, row: dict[str, Any], field_name: str, value: Any
+    ) -> None:
+        field = self.schema.model_fields.get(field_name)
+        if field is None:
+            row[field_name] = value
+        elif _is_datetime(field.annotation):
+            row[_datetime_storage_field(field_name)] = (
+                _datetime_to_ms(value) if value is not None else None
+            )
+        else:
+            row[field_name] = value
+
+    async def _query_raw(
+        self, where: Any, *, limit: int, include_vectors: bool
+    ) -> list[dict[str, Any]]:
+        await self.ensure_collection()
+        client = await get_client()
+        return await _run(
+            client.query,
+            self.collection_name,
+            filter=self._expr(where),
+            output_fields=self._output_fields(include_vectors=include_vectors),
+            limit=limit,
+        )
+
+    async def _query_candidate_rows(
+        self, where: Any, *, limit: int
+    ) -> list[dict[str, Any]]:
+        rows = await self._query_raw(where, limit=limit, include_vectors=False)
+        return [self._restore_row(row, include_distance=False) for row in rows]
+
+    def _expr(self, where: Any) -> str:
+        if where is None:
+            return ""
+        milvus_expr = getattr(where, "milvus", None)
+        if isinstance(milvus_expr, str):
+            return milvus_expr
+        if not isinstance(where, str):
+            return str(where)
+        return _lancedb_expr_to_milvus(where)
+
+
+def _lancedb_expr_to_milvus(expr: str) -> str:
+    """Best-effort conversion for internal EverOS LanceDB predicates."""
+    converted = expr
+    converted = re.sub(r"\barray_has\s*\(", "array_contains(", converted)
+    converted = re.sub(r"\bIS\s+NULL\b", "is null", converted, flags=re.IGNORECASE)
+    converted = re.sub(
+        r"\b([A-Za-z_][A-Za-z0-9_]*)\s+IN\s+\(([^)]*)\)",
+        lambda m: f"{m.group(1)} in [{m.group(2)}]",
+        converted,
+        flags=re.IGNORECASE,
+    )
+    converted = re.sub(r"(?<![!<>=])=(?!=)", "==", converted)
+    converted = re.sub(r"\btimestamp\b", "timestamp_ms", converted)
+    converted = re.sub(r"\bcreated_at\b", "created_at_ms", converted)
+    converted = re.sub(r"\bupdated_at\b", "updated_at_ms", converted)
+    converted = re.sub(
+        r"TIMESTAMP\s+'([^']*)'",
+        lambda m: str(_iso_to_ms(m.group(1))),
+        converted,
+    )
+    return converted
+
+
+def _datetime_storage_field(name: str) -> str:
+    return f"{name}_ms"
+
+
+def _sparse_field(name: str) -> str:
+    return f"{name}{_SPARSE_SUFFIX}"
+
+
+def _datetime_to_ms(value: Any) -> int:
+    if isinstance(value, dt.datetime):
+        aware = ensure_utc(value)
+        assert aware is not None
+        return to_timestamp_ms(aware)
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return int(value)
+    raise TypeError(f"expected datetime or epoch ms, got {type(value).__name__}")
+
+
+def _iso_to_ms(value: str) -> int:
+    aware = ensure_utc(from_iso_format(value))
+    assert aware is not None
+    return to_timestamp_ms(aware)
+
+
+def _is_optional(annotation: Any) -> bool:
+    args = get_args(annotation)
+    return type(None) in args
+
+
+def _non_none_args(annotation: Any) -> tuple[Any, ...]:
+    return tuple(arg for arg in get_args(annotation) if arg is not type(None))
+
+
+def _is_datetime(annotation: Any) -> bool:
+    if annotation is dt.datetime:
+        return True
+    return any(arg is dt.datetime for arg in _non_none_args(annotation))
+
+
+def _is_list(annotation: Any) -> bool:
+    origin = get_origin(annotation)
+    if origin is list:
+        return True
+    return any(get_origin(arg) is list for arg in _non_none_args(annotation))
+
+
+def _is_float(annotation: Any) -> bool:
+    if annotation is float:
+        return True
+    return any(arg is float for arg in _non_none_args(annotation))
+
+
+def _is_int(annotation: Any) -> bool:
+    if annotation is int:
+        return True
+    return any(arg is int for arg in _non_none_args(annotation))
+
+
+def _sort_value(value: Any) -> Any:
+    if value is None:
+        fallback = ensure_utc(dt.datetime.min)
+        assert fallback is not None
+        return fallback
+    return value
+
+
+def _bm25_score_from_distance(distance: Any) -> float:
+    """Convert the Milvus BM25 search distance into a higher-is-better score."""
+    if distance is None:
+        return 0.0
+    return abs(float(distance))
+
+
+def _cosine_distance_from_milvus(distance: Any) -> float | None:
+    """Normalize Milvus COSINE results to LanceDB-style distance.
+
+    Milvus server / Zilliz Cloud return cosine similarity through the
+    ``distance`` field. Milvus Lite 3.0 reports cosine distance instead; keep
+    this version-gated workaround narrow because the upstream issue is expected
+    to be fixed after 3.0:
+    https://github.com/milvus-io/milvus-lite/issues/343
+
+    Recaller code expects LanceDB's ``1 - similarity`` distance contract.
+    """
+    if distance is None:
+        return None
+    value = float(distance)
+    if not _uses_milvus_lite_3_0_cosine_distance_bug():
+        value = 1.0 - value
+    if value < 0.0:
+        return 0.0
+    if value > 1.0:
+        return 1.0
+    return value
+
+
+def _uses_milvus_lite() -> bool:
+    uri = load_settings().milvus.uri
+    if not uri:
+        return True
+    return "://" not in uri
+
+
+def _uses_milvus_lite_3_0_cosine_distance_bug() -> bool:
+    if not _uses_milvus_lite():
+        return False
+    return _milvus_lite_version() in {"3.0", "3.0.0"}
+
+
+def _milvus_lite_version() -> str | None:
+    try:
+        return version("milvus-lite")
+    except PackageNotFoundError:
+        return None
+
+
+def _first_result_set(results: Any) -> list[dict[str, Any]]:
+    if not results:
+        return []
+    first = results[0]
+    return list(first or [])
+
+
+async def _run(func: Any, /, *args: Any, **kwargs: Any) -> Any:
+    return await asyncio.to_thread(func, *args, **kwargs)
+
+
+__all__ = ["MilvusRepoBase"]
diff --git a/src/everos/memory/cascade/handlers/agent_case.py b/src/everos/memory/cascade/handlers/agent_case.py
index 453c053d5..275e30fa1 100644
--- a/src/everos/memory/cascade/handlers/agent_case.py
+++ b/src/everos/memory/cascade/handlers/agent_case.py
@@ -26,7 +26,7 @@
 
 from __future__ import annotations
 
-from everos.infra.persistence.lancedb import AgentCase, ParentType, agent_case_repo
+from everos.infra.persistence.index import AgentCase, ParentType, agent_case_repo
 
 from ._common import require_float, require_iso_timestamp
 from ._daily_log_base import BaseDailyLogHandler, ParsedEntry
diff --git a/src/everos/memory/cascade/handlers/agent_skill.py b/src/everos/memory/cascade/handlers/agent_skill.py
index 12046d526..c8831b2a2 100644
--- a/src/everos/memory/cascade/handlers/agent_skill.py
+++ b/src/everos/memory/cascade/handlers/agent_skill.py
@@ -38,7 +38,7 @@
 import anyio
 
 from everos.core.persistence import MarkdownReader
-from everos.infra.persistence.lancedb import AgentSkill, agent_skill_repo
+from everos.infra.persistence.index import AgentSkill, agent_skill_repo
 from everos.infra.persistence.markdown import AgentSkillFrontmatter
 
 from ..types import HandlerOutcome
diff --git a/src/everos/memory/cascade/handlers/atomic_fact.py b/src/everos/memory/cascade/handlers/atomic_fact.py
index a33e94ede..4389555c5 100644
--- a/src/everos/memory/cascade/handlers/atomic_fact.py
+++ b/src/everos/memory/cascade/handlers/atomic_fact.py
@@ -19,7 +19,7 @@
 
 from __future__ import annotations
 
-from everos.infra.persistence.lancedb import AtomicFact, ParentType, atomic_fact_repo
+from everos.infra.persistence.index import AtomicFact, ParentType, atomic_fact_repo
 
 from ._common import parse_inline_list, require_iso_timestamp
 from ._daily_log_base import BaseDailyLogHandler, ParsedEntry
diff --git a/src/everos/memory/cascade/handlers/episode.py b/src/everos/memory/cascade/handlers/episode.py
index e28ac8c87..4750afece 100644
--- a/src/everos/memory/cascade/handlers/episode.py
+++ b/src/everos/memory/cascade/handlers/episode.py
@@ -32,7 +32,7 @@
 
 import asyncio
 
-from everos.infra.persistence.lancedb import Episode, ParentType, episode_repo
+from everos.infra.persistence.index import Episode, ParentType, episode_repo
 
 from ._common import parse_inline_list, require_iso_timestamp
 from ._daily_log_base import BaseDailyLogHandler, ParsedEntry
diff --git a/src/everos/memory/cascade/handlers/foresight.py b/src/everos/memory/cascade/handlers/foresight.py
index 50bae7728..b9dab218e 100644
--- a/src/everos/memory/cascade/handlers/foresight.py
+++ b/src/everos/memory/cascade/handlers/foresight.py
@@ -26,7 +26,7 @@
 
 from __future__ import annotations
 
-from everos.infra.persistence.lancedb import Foresight, ParentType, foresight_repo
+from everos.infra.persistence.index import Foresight, ParentType, foresight_repo
 
 from ._common import (
     optional_int,
diff --git a/src/everos/memory/cascade/handlers/knowledge_topic.py b/src/everos/memory/cascade/handlers/knowledge_topic.py
index 80e2cdfd4..254124ed8 100644
--- a/src/everos/memory/cascade/handlers/knowledge_topic.py
+++ b/src/everos/memory/cascade/handlers/knowledge_topic.py
@@ -34,7 +34,7 @@
 
 from everos.component.utils.datetime import get_utc_now
 from everos.core.persistence import MarkdownReader, ParsedMarkdown
-from everos.infra.persistence.lancedb import KnowledgeTopic, knowledge_topic_repo
+from everos.infra.persistence.index import KnowledgeTopic, knowledge_topic_repo
 from everos.infra.persistence.sqlite import (
     TopicUpsertPayload,
     knowledge_topic_sqlite_repo,
diff --git a/src/everos/memory/cascade/handlers/user_profile.py b/src/everos/memory/cascade/handlers/user_profile.py
index 4842986d0..8d7573521 100644
--- a/src/everos/memory/cascade/handlers/user_profile.py
+++ b/src/everos/memory/cascade/handlers/user_profile.py
@@ -27,7 +27,7 @@
 from typing import Any, ClassVar
 
 from everos.core.persistence import MarkdownReader
-from everos.infra.persistence.lancedb import UserProfile, user_profile_repo
+from everos.infra.persistence.index import UserProfile, user_profile_repo
 
 from ..types import HandlerOutcome
 from ._common import content_sha256 as compute_content_sha256
diff --git a/src/everos/memory/cascade/registry.py b/src/everos/memory/cascade/registry.py
index 34de209e3..e74162017 100644
--- a/src/everos/memory/cascade/registry.py
+++ b/src/everos/memory/cascade/registry.py
@@ -19,7 +19,7 @@
 from pathlib import PurePosixPath
 
 from everos.core.persistence.markdown import BaseFrontmatter
-from everos.infra.persistence.lancedb import (
+from everos.infra.persistence.index import (
     AgentCase,
     AgentSkill,
     AtomicFact,
diff --git a/src/everos/memory/get/filters_adapter.py b/src/everos/memory/get/filters_adapter.py
index 407a06bc7..46b1e6305 100644
--- a/src/everos/memory/get/filters_adapter.py
+++ b/src/everos/memory/get/filters_adapter.py
@@ -16,6 +16,7 @@
 from __future__ import annotations
 
 from everos.memory.search import FilterNode, compile_filters
+from everos.memory.search.filters import BackendFilters, compile_filters_for_backends
 
 
 def compile_filters_for_get(
@@ -38,3 +39,21 @@ def compile_filters_for_get(
         app_id=app_id,
         project_id=project_id,
     )
+
+
+def compile_filters_for_get_backends(
+    filters: FilterNode | None,
+    *,
+    owner_id: str,
+    owner_type: str,
+    app_id: str = "default",
+    project_id: str = "default",
+) -> BackendFilters:
+    """Compile ``/get`` filters for every derived index backend."""
+    return compile_filters_for_backends(
+        filters,
+        owner_id=owner_id,
+        owner_type=owner_type,
+        app_id=app_id,
+        project_id=project_id,
+    )
diff --git a/src/everos/memory/get/manager.py b/src/everos/memory/get/manager.py
index 1000cbdd2..e4ddaa7f8 100644
--- a/src/everos/memory/get/manager.py
+++ b/src/everos/memory/get/manager.py
@@ -12,7 +12,7 @@
 Reads only — never writes. Filters are compiled through
 :func:`compile_filters_for_get` so the column allow-list stays
 shared with :mod:`memory.search`. Pagination + in-memory sort
-runs through :meth:`LanceRepoBase.find_where_paginated`.
+runs through the configured derived index repository.
 """
 
 from __future__ import annotations
@@ -34,11 +34,11 @@
     GetRequest,
     GetResponse,
 )
-from .filters_adapter import compile_filters_for_get
+from .filters_adapter import compile_filters_for_get_backends
 
 if TYPE_CHECKING:
     from everos.core.persistence.lancedb import LanceRepoBase
-    from everos.infra.persistence.lancedb import (
+    from everos.infra.persistence.index import (
         AgentCase,
         AgentSkill,
         Episode,
@@ -49,7 +49,7 @@
 
 
 class GetManager:
-    """Dispatch ``GetRequest`` to the matching LanceDB-backed repo and
+    """Dispatch ``GetRequest`` to the matching derived index repo and
     shape rows into the public DTO."""
 
     def __init__(
@@ -70,7 +70,7 @@ def __init__(
     async def get(self, req: GetRequest) -> GetResponse:
         request_id = gen_request_id()
         descending = req.sort_order == "desc"
-        where = compile_filters_for_get(
+        where = compile_filters_for_get_backends(
             req.filters,
             owner_id=req.owner_id,
             owner_type=req.owner_type,
@@ -187,7 +187,7 @@ def _shape_agent_skill(row: AgentSkill) -> GetAgentSkillItem:
 
     async def _fetch_profile(self, owner_id: str) -> list[GetProfileItem]:
         """Fetch the owner's single profile row from the ``user_profile``
-        LanceDB table (kept in sync with ``users/<id>/user.md`` by cascade).
+        Derived index row (kept in sync with ``users/<id>/user.md`` by cascade).
 
         Profile is one-row-per-owner KV — there is no pagination / sort /
         filter surface, so at most one item is returned. Mirrors the
diff --git a/src/everos/memory/search/filters.py b/src/everos/memory/search/filters.py
index 2dd9aa804..0f417afc0 100644
--- a/src/everos/memory/search/filters.py
+++ b/src/everos/memory/search/filters.py
@@ -1,4 +1,4 @@
-"""Filters DSL → LanceDB ``where`` string compiler.
+"""Filters DSL → derived-index filter compiler.
 
 The Filters DSL is intentionally permissive at the JSON layer (so callers
 can pass whatever they like and get a clean 400 if it is not supported)
@@ -23,7 +23,8 @@
   kind). Iterate / membership-test only; do not mutate.
 * :data:`RESERVED_FIELDS` — names rejected inside any ``filters`` block.
 * :func:`compile_predicate` — render one ``{field: value}`` clause to
-  SQL. Operator-map and equality-shorthand are both handled.
+  a backend-specific predicate. Operator-map and equality-shorthand are
+  both handled.
 
 The high-level :func:`compile_filters` remains the entry point for
 ``/search`` (combinator-aware).
@@ -31,10 +32,18 @@
 
 from __future__ import annotations
 
+import dataclasses
 import datetime as _dt
+import json
 from typing import Any, Final
 
-from everos.component.utils.datetime import from_timestamp, to_iso_format
+from everos.component.utils.datetime import (
+    ensure_utc,
+    from_iso_format,
+    from_timestamp,
+    to_iso_format,
+    to_timestamp_ms,
+)
 from everos.core.errors import FilterError as FilterError
 
 from .dto import FilterNode
@@ -50,6 +59,15 @@
     "lte": "<=",
     "in": "IN",
 }
+_MILVUS_OP_MAP: Final[dict[str, str]] = {
+    "eq": "==",
+    "ne": "!=",
+    "gt": ">",
+    "gte": ">=",
+    "lt": "<",
+    "lte": "<=",
+    "in": "in",
+}
 
 # Field kinds: ``str`` rendered as ``'<escaped>'``; ``ts`` rendered as
 # ``TIMESTAMP '<iso>'`` (DataFusion timestamp literal); ``array_str``
@@ -82,6 +100,17 @@ def __init__(self, column: str, kind: _FieldKind) -> None:
 )
 
 
+@dataclasses.dataclass(frozen=True)
+class BackendFilters:
+    """Compiled filters for every supported derived index backend."""
+
+    lancedb: str
+    milvus: str
+
+    def __str__(self) -> str:
+        return self.lancedb
+
+
 # ── Public API ───────────────────────────────────────────────────────────
 
 
@@ -102,19 +131,70 @@ def compile_filters(
     from another — omitting it would let a query bleed across spaces. Both
     ``/search`` and ``/get`` share this compile path.
     """
-    base = [
-        f"owner_id = '{_escape_str(owner_id)}'",
-        f"owner_type = '{owner_type}'",
-        f"app_id = '{_escape_str(app_id)}'",
-        f"project_id = '{_escape_str(project_id)}'",
-    ]
+    return _compile_filters_backend(
+        node,
+        owner_id=owner_id,
+        owner_type=owner_type,
+        app_id=app_id,
+        project_id=project_id,
+        backend="lancedb",
+    )
+
+
+def compile_filters_for_backends(
+    node: FilterNode | None,
+    *,
+    owner_id: str,
+    owner_type: str,
+    app_id: str = "default",
+    project_id: str = "default",
+) -> BackendFilters:
+    """Compile request filters for every supported derived index backend."""
+    return BackendFilters(
+        lancedb=_compile_filters_backend(
+            node,
+            owner_id=owner_id,
+            owner_type=owner_type,
+            app_id=app_id,
+            project_id=project_id,
+            backend="lancedb",
+        ),
+        milvus=_compile_filters_backend(
+            node,
+            owner_id=owner_id,
+            owner_type=owner_type,
+            app_id=app_id,
+            project_id=project_id,
+            backend="milvus",
+        ),
+    )
+
+
+def _compile_filters_backend(
+    node: FilterNode | None,
+    *,
+    owner_id: str,
+    owner_type: str,
+    app_id: str,
+    project_id: str,
+    backend: str,
+) -> str:
+    base = _base_clauses(
+        owner_id=owner_id,
+        owner_type=owner_type,
+        app_id=app_id,
+        project_id=project_id,
+        backend=backend,
+    )
     # Only episode / atomic_fact tables carry the ``deprecated_by`` column
     # (Reflection V1 marks superseded entries). Agent tables don't have it.
     if owner_type == "user":
-        base.append("deprecated_by IS NULL")
+        base.append(
+            "deprecated_by IS NULL" if backend == "lancedb" else "deprecated_by is null"
+        )
     if node is None:
         return " AND ".join(base)
-    compiled = _compile_node(node.model_dump(exclude_none=True))
+    compiled = _compile_node(node.model_dump(exclude_none=True), backend=backend)
     if not compiled:
         return " AND ".join(base)
     return " AND ".join([*base, compiled])
@@ -123,7 +203,7 @@ def compile_filters(
 # ── Internals ────────────────────────────────────────────────────────────
 
 
-def _compile_node(raw: dict[str, Any]) -> str:
+def _compile_node(raw: dict[str, Any], *, backend: str = "lancedb") -> str:
     """Walk one DSL node; return the matching SQL fragment (no leading parens).
 
     Empty nodes yield ``""`` so :func:`compile_filters` can skip the
@@ -133,9 +213,9 @@ def _compile_node(raw: dict[str, Any]) -> str:
     parts: list[str] = []
 
     if (and_list := raw.pop("AND", None)) is not None:
-        parts.append(_compile_combinator(and_list, "AND"))
+        parts.append(_compile_combinator(and_list, "AND", backend=backend))
     if (or_list := raw.pop("OR", None)) is not None:
-        parts.append(_compile_combinator(or_list, "OR"))
+        parts.append(_compile_combinator(or_list, "OR", backend=backend))
 
     for field, value in raw.items():
         if field in RESERVED_FIELDS:
@@ -144,7 +224,7 @@ def _compile_node(raw: dict[str, Any]) -> str:
             )
         if field not in ALLOWED_FIELDS:
             raise FilterError(f"unsupported filter field: {field!r}")
-        parts.append(compile_predicate(field, value))
+        parts.append(compile_predicate(field, value, backend=backend))
 
     # Drop empty fragments coming from empty AND/OR arrays.
     parts = [p for p in parts if p]
@@ -155,7 +235,9 @@ def _compile_node(raw: dict[str, Any]) -> str:
     return " AND ".join(parts)
 
 
-def _compile_combinator(children: list[dict[str, Any]], op: str) -> str:
+def _compile_combinator(
+    children: list[dict[str, Any]], op: str, *, backend: str = "lancedb"
+) -> str:
     """Render an ``AND`` / ``OR`` array of child nodes."""
     if not isinstance(children, list):
         raise FilterError(f"{op} expects an array of nodes")
@@ -163,7 +245,7 @@ def _compile_combinator(children: list[dict[str, Any]], op: str) -> str:
     for child in children:
         if not isinstance(child, dict):
             raise FilterError(f"{op} children must be objects")
-        compiled = _compile_node(child)
+        compiled = _compile_node(child, backend=backend)
         if compiled:
             fragments.append(f"({compiled})")
     if not fragments:
@@ -175,7 +257,7 @@ def _compile_combinator(children: list[dict[str, Any]], op: str) -> str:
     return "(" + glue.join(fragments) + ")"
 
 
-def compile_predicate(field: str, value: Any) -> str:
+def compile_predicate(field: str, value: Any, *, backend: str = "lancedb") -> str:
     """Render one ``"<field>": <value>`` clause to SQL.
 
     Public primitive — :mod:`memory.get` builds a flat (no AND/OR)
@@ -195,48 +277,65 @@ def compile_predicate(field: str, value: Any) -> str:
         if not value:
             raise FilterError(f"empty operator map for field {field!r}")
         clauses = [
-            _compile_op_clause(spec, field, op, op_val) for op, op_val in value.items()
+            _compile_op_clause(spec, field, op, op_val, backend=backend)
+            for op, op_val in value.items()
         ]
         if len(clauses) == 1:
             return clauses[0]
         return "(" + " AND ".join(clauses) + ")"
     # Equality shorthand.
-    return _compile_op_clause(spec, field, "eq", value)
+    return _compile_op_clause(spec, field, "eq", value, backend=backend)
 
 
-def _compile_op_clause(spec: _FieldSpec, field: str, op: str, value: Any) -> str:
+def _compile_op_clause(
+    spec: _FieldSpec, field: str, op: str, value: Any, *, backend: str = "lancedb"
+) -> str:
     """Render a single ``<field> <op> <value>`` clause."""
     if op not in _OP_MAP:
         raise FilterError(f"unsupported operator {op!r} on field {field!r}")
-    sql_op = _OP_MAP[op]
+    op_map = _MILVUS_OP_MAP if backend == "milvus" else _OP_MAP
+    sql_op = op_map[op]
+    column = _column_for_backend(spec, backend)
 
     if spec.kind == "array_str":
         # Only equality / membership make sense on a list column.
+        fn = "array_contains" if backend == "milvus" else "array_has"
         if op == "eq":
-            literal = _escape_str(_require_str(value, field))
-            return f"array_has({spec.column}, '{literal}')"
+            literal = _render_str_literal(_require_str(value, field), backend)
+            return f"{fn}({column}, {literal})"
         if op == "in":
             items = _require_list(value, field)
-            literals = [f"'{_escape_str(_require_str(v, field))}'" for v in items]
-            inner = " OR ".join(f"array_has({spec.column}, {lit})" for lit in literals)
+            literals = [
+                _render_str_literal(_require_str(v, field), backend) for v in items
+            ]
+            inner = " OR ".join(f"{fn}({column}, {lit})" for lit in literals)
             return f"({inner})"
         raise FilterError(f"operator {op!r} is not supported on array field {field!r}")
 
     if op == "in":
         items = _require_list(value, field)
-        literals = [_render_literal(v, spec.kind, field) for v in items]
-        return f"{spec.column} IN ({', '.join(literals)})"
+        literals = [
+            _render_literal(v, spec.kind, field, backend=backend) for v in items
+        ]
+        if backend == "milvus":
+            return f"{column} in [{', '.join(literals)}]"
+        return f"{column} IN ({', '.join(literals)})"
 
-    return f"{spec.column} {sql_op} {_render_literal(value, spec.kind, field)}"
+    literal = _render_literal(value, spec.kind, field, backend=backend)
+    return f"{column} {sql_op} {literal}"
 
 
 # ── Literal rendering ────────────────────────────────────────────────────
 
 
-def _render_literal(value: Any, kind: _FieldKind, field: str) -> str:
+def _render_literal(
+    value: Any, kind: _FieldKind, field: str, *, backend: str = "lancedb"
+) -> str:
     if kind == "str":
-        return f"'{_escape_str(_require_str(value, field))}'"
+        return _render_str_literal(_require_str(value, field), backend)
     if kind == "ts":
+        if backend == "milvus":
+            return str(_render_ts_ms(value, field))
         return f"TIMESTAMP '{_render_ts(value, field)}'"
     raise FilterError(f"unsupported field kind {kind!r} for field {field!r}")
 
@@ -257,6 +356,58 @@ def _render_ts(value: Any, field: str) -> str:
     raise FilterError(f"timestamp value for {field!r} must be ms or ISO string")
 
 
+def _render_ts_ms(value: Any, field: str) -> int:
+    if isinstance(value, bool):
+        raise FilterError(f"timestamp value for {field!r} must be ms or ISO string")
+    if isinstance(value, (int, float)):
+        return int(value)
+    if isinstance(value, str):
+        if "'" in value:
+            raise FilterError(f"timestamp string for {field!r} contains a quote")
+        try:
+            aware = ensure_utc(from_iso_format(value))
+        except (TypeError, ValueError) as exc:
+            raise FilterError(
+                f"timestamp value for {field!r} must be ms or ISO string"
+            ) from exc
+        assert aware is not None
+        return to_timestamp_ms(aware)
+    if isinstance(value, _dt.datetime):
+        aware = ensure_utc(value)
+        assert aware is not None
+        return to_timestamp_ms(aware)
+    raise FilterError(f"timestamp value for {field!r} must be ms or ISO string")
+
+
+def _render_str_literal(value: str, backend: str) -> str:
+    if backend == "milvus":
+        return json.dumps(value, ensure_ascii=False)
+    return f"'{_escape_str(value)}'"
+
+
+def _base_clauses(
+    *,
+    owner_id: str,
+    owner_type: str,
+    app_id: str,
+    project_id: str,
+    backend: str,
+) -> list[str]:
+    eq = "==" if backend == "milvus" else "="
+    return [
+        f"owner_id {eq} {_render_str_literal(owner_id, backend)}",
+        f"owner_type {eq} {_render_str_literal(owner_type, backend)}",
+        f"app_id {eq} {_render_str_literal(app_id, backend)}",
+        f"project_id {eq} {_render_str_literal(project_id, backend)}",
+    ]
+
+
+def _column_for_backend(spec: _FieldSpec, backend: str) -> str:
+    if backend == "milvus" and spec.kind == "ts":
+        return f"{spec.column}_ms"
+    return spec.column
+
+
 def _escape_str(value: str) -> str:
     """Double single quotes — SQL-standard escape for a single-quoted literal."""
     return value.replace("'", "''")
diff --git a/src/everos/memory/search/manager.py b/src/everos/memory/search/manager.py
index 7633df415..e86672c7b 100644
--- a/src/everos/memory/search/manager.py
+++ b/src/everos/memory/search/manager.py
@@ -22,7 +22,8 @@
 no query-relevance score we can assign to a fact pulled by parent_id
 alone, and emitting ``score=0.0`` facts would muddy the contract.
 
-The manager never writes to storage; it only reads LanceDB + markdown.
+The manager never writes to storage; it only reads the derived index +
+markdown.
 """
 
 from __future__ import annotations
@@ -59,7 +60,7 @@
     SearchResponse,
     UnprocessedMessageDTO,
 )
-from .filters import compile_filters
+from .filters import compile_filters_for_backends
 from .hierarchy import build_ep_to_fact_parents, heap_expand
 from .shaper import (
     reshape_hybrid_output,
@@ -156,7 +157,7 @@ async def search(self, req: SearchRequest) -> SearchResponse:
         # Compile filters first: a malformed `filters` payload is a user
         # input error (422) and should surface before the server-side
         # component guard (500). The two steps are independent.
-        where = compile_filters(
+        where = compile_filters_for_backends(
             req.filters,
             owner_id=req.owner_id,
             owner_type=req.owner_type,
diff --git a/src/everos/memory/search/recall/agent_case.py b/src/everos/memory/search/recall/agent_case.py
index 93a7f9b60..9c17ba45e 100644
--- a/src/everos/memory/search/recall/agent_case.py
+++ b/src/everos/memory/search/recall/agent_case.py
@@ -13,17 +13,15 @@
 
 from __future__ import annotations
 
-import asyncio
 from collections.abc import Sequence
 from typing import ClassVar
 
 from everalgo.types import Candidate
 
-from everos.infra.persistence.lancedb import AgentCase, get_table
+from everos.infra.persistence.index import AgentCase, agent_case_repo
 
 from .base import (
     RecallerDeps,
-    build_or_query_multi_column,
     cosine_score_from_distance,
     row_to_candidate,
 )
@@ -50,43 +48,15 @@ async def sparse_recall(
         per BM25 column (``MatchQuery`` is column-bound), then the
         two per-column result lists merge by id keeping the max score.
         """
-        column_queries = build_or_query_multi_column(
-            self._deps.tokenizer, query, AgentCase.BM25_FIELDS
-        )
-        if column_queries is None:
+        terms = [term for term in self._deps.tokenizer.tokenize(query) if term]
+        if not terms:
             return []
-        table = await get_table(AgentCase.TABLE_NAME, AgentCase)
-
-        async def _query_one(column: str) -> list[dict]:
-            return (
-                await table.query()
-                .nearest_to_text(column_queries[column])
-                .where(where)
-                .limit(limit)
-                .to_list()
-            )
-
-        per_column = await asyncio.gather(
-            *(_query_one(col) for col in AgentCase.BM25_FIELDS),
+        merged_rows = await agent_case_repo.sparse_search(
+            terms,
+            where,
+            columns=AgentCase.BM25_FIELDS,
+            limit=limit,
         )
-        # Merge by id, keep the max BM25 score across the two columns.
-        # task_intent hits typically score higher (the retrieval anchor);
-        # approach hits catch queries that match a step detail.
-        best: dict[str, dict] = {}
-        for rows in per_column:
-            for r in rows:
-                rid = r.get("id")
-                if not isinstance(rid, str):
-                    continue
-                score = float(r.get("_score", 0.0))
-                existing = best.get(rid)
-                if existing is None or score > float(existing.get("_score", 0.0)):
-                    merged = dict(r)
-                    merged["_score"] = score
-                    best[rid] = merged
-        merged_rows = sorted(
-            best.values(), key=lambda r: float(r.get("_score", 0.0)), reverse=True
-        )[:limit]
         return [
             row_to_candidate(r, source="keyword", score=float(r.get("_score", 0.0)))
             for r in merged_rows
@@ -97,15 +67,7 @@ async def dense_recall(
     ) -> list[Candidate]:
         if not vector:
             return []
-        table = await get_table(AgentCase.TABLE_NAME, AgentCase)
-        rows = (
-            await table.query()
-            .nearest_to(list(vector))
-            .distance_type("cosine")
-            .where(where)
-            .limit(limit)
-            .to_list()
-        )
+        rows = await agent_case_repo.dense_search(vector, where, limit=limit)
         return [
             row_to_candidate(
                 r,
diff --git a/src/everos/memory/search/recall/agent_skill.py b/src/everos/memory/search/recall/agent_skill.py
index f444e4a38..0c1831c79 100644
--- a/src/everos/memory/search/recall/agent_skill.py
+++ b/src/everos/memory/search/recall/agent_skill.py
@@ -9,17 +9,15 @@
 
 from __future__ import annotations
 
-import asyncio
 from collections.abc import Sequence
 from typing import ClassVar
 
 from everalgo.types import Candidate
 
-from everos.infra.persistence.lancedb import AgentSkill, get_table
+from everos.infra.persistence.index import AgentSkill, agent_skill_repo
 
 from .base import (
     RecallerDeps,
-    build_or_query_multi_column,
     cosine_score_from_distance,
     row_to_candidate,
 )
@@ -48,41 +46,15 @@ async def sparse_recall(
         rationale. One BooleanQuery per BM25 column; merge by id with
         max score.
         """
-        column_queries = build_or_query_multi_column(
-            self._deps.tokenizer, query, AgentSkill.BM25_FIELDS
-        )
-        if column_queries is None:
+        terms = [term for term in self._deps.tokenizer.tokenize(query) if term]
+        if not terms:
             return []
-        table = await get_table(AgentSkill.TABLE_NAME, AgentSkill)
-
-        async def _query_one(column: str) -> list[dict]:
-            return (
-                await table.query()
-                .nearest_to_text(column_queries[column])
-                .where(where)
-                .limit(limit)
-                .to_list()
-            )
-
-        per_column = await asyncio.gather(
-            *(_query_one(col) for col in AgentSkill.BM25_FIELDS),
+        merged_rows = await agent_skill_repo.sparse_search(
+            terms,
+            where,
+            columns=AgentSkill.BM25_FIELDS,
+            limit=limit,
         )
-        # Merge by id, keep max BM25 score across the two columns.
-        best: dict[str, dict] = {}
-        for rows in per_column:
-            for r in rows:
-                rid = r.get("id")
-                if not isinstance(rid, str):
-                    continue
-                score = float(r.get("_score", 0.0))
-                existing = best.get(rid)
-                if existing is None or score > float(existing.get("_score", 0.0)):
-                    merged = dict(r)
-                    merged["_score"] = score
-                    best[rid] = merged
-        merged_rows = sorted(
-            best.values(), key=lambda r: float(r.get("_score", 0.0)), reverse=True
-        )[:limit]
         return [
             row_to_candidate(r, source="keyword", score=float(r.get("_score", 0.0)))
             for r in merged_rows
@@ -93,15 +65,7 @@ async def dense_recall(
     ) -> list[Candidate]:
         if not vector:
             return []
-        table = await get_table(AgentSkill.TABLE_NAME, AgentSkill)
-        rows = (
-            await table.query()
-            .nearest_to(list(vector))
-            .distance_type("cosine")
-            .where(where)
-            .limit(limit)
-            .to_list()
-        )
+        rows = await agent_skill_repo.dense_search(vector, where, limit=limit)
         return [
             row_to_candidate(
                 r,
@@ -124,8 +88,7 @@ async def fetch_by_case_ids(
         """
         if not case_ids:
             return []
-        table = await get_table(AgentSkill.TABLE_NAME, AgentSkill)
         clause = " OR ".join(f"array_has(source_case_ids, '{_q(c)}')" for c in case_ids)
         full_where = f"({where}) AND ({clause})"
-        rows = await table.query().where(full_where).limit(limit).to_list()
+        rows = await agent_skill_repo.search(where=full_where, limit=limit)
         return [row_to_candidate(r, source="vector", score=0.0) for r in rows]
diff --git a/src/everos/memory/search/recall/atomic_fact.py b/src/everos/memory/search/recall/atomic_fact.py
index 69eba22d4..f78b3f50a 100644
--- a/src/everos/memory/search/recall/atomic_fact.py
+++ b/src/everos/memory/search/recall/atomic_fact.py
@@ -21,11 +21,10 @@
 
 from everalgo.types import Candidate, FactCandidate
 
-from everos.infra.persistence.lancedb import AtomicFact, get_table
+from everos.infra.persistence.index import AtomicFact, atomic_fact_repo
 
 from .base import (
     RecallerDeps,
-    build_or_query,
     cosine_score_from_distance,
     row_to_candidate,
 )
@@ -33,6 +32,7 @@
 _NOISE_COLUMNS = frozenset(
     {"vector", "_distance", "_score", "created_at", "updated_at"}
 )
+_MAX_FACT_RECALL_LIMIT = 1024
 
 
 class AtomicFactRecaller:
@@ -49,14 +49,14 @@ async def sparse_recall(
         self, query: str, where: str, *, limit: int
     ) -> list[Candidate]:
         """BM25 recall via OR-mode BooleanQuery (see EpisodeRecaller docstring)."""
-        bq = build_or_query(
-            self._deps.tokenizer, query, column=AtomicFact.BM25_FIELDS[0]
-        )
-        if bq is None:
+        terms = [term for term in self._deps.tokenizer.tokenize(query) if term]
+        if not terms:
             return []
-        table = await get_table(AtomicFact.TABLE_NAME, AtomicFact)
-        rows = (
-            await table.query().nearest_to_text(bq).where(where).limit(limit).to_list()
+        rows = await atomic_fact_repo.sparse_search(
+            terms,
+            where,
+            columns=AtomicFact.BM25_FIELDS,
+            limit=limit,
         )
         return [
             row_to_candidate(r, source="keyword", score=float(r.get("_score", 0.0)))
@@ -78,15 +78,7 @@ async def dense_recall(
         """
         if not vector:
             return []
-        table = await get_table(AtomicFact.TABLE_NAME, AtomicFact)
-        rows = (
-            await table.query()
-            .nearest_to(list(vector))
-            .distance_type("cosine")
-            .where(where)
-            .limit(limit)
-            .to_list()
-        )
+        rows = await atomic_fact_repo.dense_search(vector, where, limit=limit)
         return [
             row_to_candidate(
                 r,
@@ -169,18 +161,15 @@ async def _query_facts_for_parents(
         quoted = ", ".join(f"'{_q(pid)}'" for pid in parent_to_eps)
         clause = f"parent_id IN ({quoted})"
         full_where = f"({where}) AND ({clause})"
-        limit = per_episode * max(len(parent_to_eps), 1)
-        table = await get_table(AtomicFact.TABLE_NAME, AtomicFact)
+        # Milvus server / Zilliz Cloud reject search topK values above 1024.
+        # The fact expansion only needs a bounded candidate pool for top-N
+        # competition, so keep the same cap for every derived-index backend.
+        limit = min(per_episode * max(len(parent_to_eps), 1), _MAX_FACT_RECALL_LIMIT)
         if query_vector:
-            return await (
-                table.query()
-                .nearest_to(list(query_vector))
-                .distance_type("cosine")
-                .where(full_where)
-                .limit(limit)
-                .to_list()
+            return await atomic_fact_repo.dense_search(
+                query_vector, full_where, limit=limit
             )
-        return await table.query().where(full_where).limit(limit).to_list()
+        return await atomic_fact_repo.search(where=full_where, limit=limit)
 
 
 def _build_parent_to_episode_map(
diff --git a/src/everos/memory/search/recall/episode.py b/src/everos/memory/search/recall/episode.py
index a2766dcec..1f2c13821 100644
--- a/src/everos/memory/search/recall/episode.py
+++ b/src/everos/memory/search/recall/episode.py
@@ -7,11 +7,10 @@
 
 from everalgo.types import Candidate
 
-from everos.infra.persistence.lancedb import Episode, get_table
+from everos.infra.persistence.index import Episode, active_backend, episode_repo
 
 from .base import (
     RecallerDeps,
-    build_or_query,
     cosine_score_from_distance,
     row_to_candidate,
 )
@@ -55,12 +54,14 @@ async def sparse_recall(
         Mirrors enterprise's ``bool.should + minimum_should_match=1``
         ES design.
         """
-        bq = build_or_query(self._deps.tokenizer, query, column=Episode.BM25_FIELDS[0])
-        if bq is None:
+        terms = [term for term in self._deps.tokenizer.tokenize(query) if term]
+        if not terms:
             return []
-        table = await get_table(Episode.TABLE_NAME, Episode)
-        rows = (
-            await table.query().nearest_to_text(bq).where(where).limit(limit).to_list()
+        rows = await episode_repo.sparse_search(
+            terms,
+            where,
+            columns=Episode.BM25_FIELDS,
+            limit=limit,
         )
         return [
             row_to_candidate(r, source="keyword", score=float(r.get("_score", 0.0)))
@@ -72,16 +73,7 @@ async def dense_recall(
     ) -> list[Candidate]:
         if not vector:
             return []
-        table = await get_table(Episode.TABLE_NAME, Episode)
-        rows = (
-            await table.query()
-            .nearest_to(list(vector))
-            .column("vector")
-            .distance_type("cosine")
-            .where(where)
-            .limit(limit)
-            .to_list()
-        )
+        rows = await episode_repo.dense_search(vector, where, limit=limit)
         return [
             row_to_candidate(
                 r,
@@ -113,6 +105,10 @@ async def dense_recall_subject(
         """
         if not vector:
             return []
+        if active_backend() == "milvus":
+            return await self.dense_recall(vector, where, limit=limit)
+        from everos.infra.persistence.lancedb import get_table
+
         table = await get_table(Episode.TABLE_NAME, Episode)
         rows = (
             await table.query()
@@ -149,8 +145,7 @@ async def fetch_all_for_owner(self, where: str) -> list[Candidate]:
         No ``limit`` — the full owner partition is required for cluster
         membership matching.
         """
-        table = await get_table(Episode.TABLE_NAME, Episode)
-        rows = await table.query().where(where).to_list()
+        rows = await episode_repo.search(where=where, limit=10_000)
         result: list[Candidate] = []
         for r in rows:
             entry_id = r.get("entry_id")
@@ -173,8 +168,7 @@ async def fetch_by_entry_ids(
         """Fetch episodes by entry_id (for facts whose parent_id is an entry_id)."""
         if not entry_ids:
             return []
-        table = await get_table(Episode.TABLE_NAME, Episode)
         quoted = ", ".join(f"'{_q(eid)}'" for eid in entry_ids)
         full_where = f"({where}) AND (entry_id IN ({quoted}))"
-        rows = await table.query().where(full_where).limit(len(entry_ids)).to_list()
+        rows = await episode_repo.search(where=full_where, limit=len(entry_ids))
         return [row_to_candidate(r, source="vector", score=0.0) for r in rows]
diff --git a/src/everos/memory/search/recall/knowledge_topic.py b/src/everos/memory/search/recall/knowledge_topic.py
index ea484e17a..9cc6a9bfb 100644
--- a/src/everos/memory/search/recall/knowledge_topic.py
+++ b/src/everos/memory/search/recall/knowledge_topic.py
@@ -12,17 +12,15 @@
 
 from __future__ import annotations
 
-import asyncio
 from collections.abc import Sequence
 from typing import ClassVar
 
 from everalgo.types import Candidate
 
-from everos.infra.persistence.lancedb import KnowledgeTopic, get_table
+from everos.infra.persistence.index import KnowledgeTopic, knowledge_topic_repo
 
 from .base import (
     RecallerDeps,
-    build_or_query_multi_column,
     cosine_score_from_distance,
     row_to_candidate,
 )
@@ -78,26 +76,15 @@ async def sparse_recall(
         matching the query in either its summary or its content body is
         surfaced without double-counting.
         """
-        column_queries = build_or_query_multi_column(
-            self._deps.tokenizer, query, KnowledgeTopic.BM25_FIELDS
-        )
-        if column_queries is None:
+        terms = [term for term in self._deps.tokenizer.tokenize(query) if term]
+        if not terms:
             return []
-        table = await get_table(KnowledgeTopic.TABLE_NAME, KnowledgeTopic)
-
-        async def _query_one(column: str) -> list[dict]:
-            return (
-                await table.query()
-                .nearest_to_text(column_queries[column])
-                .where(where)
-                .limit(limit)
-                .to_list()
-            )
-
-        per_column = await asyncio.gather(
-            *(_query_one(col) for col in KnowledgeTopic.BM25_FIELDS),
+        merged_rows = await knowledge_topic_repo.sparse_search(
+            terms,
+            where,
+            columns=KnowledgeTopic.BM25_FIELDS,
+            limit=limit,
         )
-        merged_rows = _merge_bm25_results(per_column, limit=limit)
         return [
             row_to_candidate(r, source="keyword", score=float(r.get("_score", 0.0)))
             for r in merged_rows
@@ -109,15 +96,7 @@ async def dense_recall(
         """Cosine ANN over the ``summary`` vector (1024-d)."""
         if not vector:
             return []
-        table = await get_table(KnowledgeTopic.TABLE_NAME, KnowledgeTopic)
-        rows = (
-            await table.query()
-            .nearest_to(list(vector))
-            .distance_type("cosine")
-            .where(where)
-            .limit(limit)
-            .to_list()
-        )
+        rows = await knowledge_topic_repo.dense_search(vector, where, limit=limit)
         return [
             row_to_candidate(
                 r,
diff --git a/src/everos/memory/search/recall/profile.py b/src/everos/memory/search/recall/profile.py
index be149b927..640d27091 100644
--- a/src/everos/memory/search/recall/profile.py
+++ b/src/everos/memory/search/recall/profile.py
@@ -19,7 +19,7 @@
 from typing import Any
 
 from everos.core.observability.logging import get_logger
-from everos.infra.persistence.lancedb import user_profile_repo
+from everos.infra.persistence.index import user_profile_repo
 
 from ..dto import SearchProfileItem
 
diff --git a/src/everos/memory/strategies/extract_agent_skill.py b/src/everos/memory/strategies/extract_agent_skill.py
index ceccf5dfb..b9d5cb164 100644
--- a/src/everos/memory/strategies/extract_agent_skill.py
+++ b/src/everos/memory/strategies/extract_agent_skill.py
@@ -49,13 +49,13 @@
 from everos.infra.ome.context import StrategyContext
 from everos.infra.ome.decorator import offline_strategy
 from everos.infra.ome.triggers import Immediate
-from everos.infra.persistence.lancedb import (
+from everos.infra.persistence.index import (
     AgentCase as LanceAgentCase,
 )
-from everos.infra.persistence.lancedb import (
+from everos.infra.persistence.index import (
     AgentSkill as LanceAgentSkill,
 )
-from everos.infra.persistence.lancedb import (
+from everos.infra.persistence.index import (
     agent_case_repo,
     agent_skill_repo,
 )
@@ -205,7 +205,7 @@ async def _load_target_case(
     if target is None:
         # Cascade hasn't indexed the freshly-written md yet.
         raise _CaseNotYetIndexedError(
-            f"AgentCase entry_id={case_entry_id} not in LanceDB yet; retrying"
+            f"AgentCase entry_id={case_entry_id} not in derived index yet; retrying"
         )
     return target
 
diff --git a/src/everos/memory/strategies/extract_user_profile.py b/src/everos/memory/strategies/extract_user_profile.py
index c4c9a6e29..7cc064548 100644
--- a/src/everos/memory/strategies/extract_user_profile.py
+++ b/src/everos/memory/strategies/extract_user_profile.py
@@ -37,7 +37,7 @@
 from everos.infra.ome.context import StrategyContext
 from everos.infra.ome.decorator import offline_strategy
 from everos.infra.ome.triggers import Immediate
-from everos.infra.persistence.lancedb import episode_repo
+from everos.infra.persistence.index import episode_repo
 from everos.infra.persistence.markdown import (
     ProfileReader,
     ProfileWriter,
diff --git a/src/everos/memory/strategies/reflect_episodes.py b/src/everos/memory/strategies/reflect_episodes.py
index 8f72f2ac3..5a208e822 100644
--- a/src/everos/memory/strategies/reflect_episodes.py
+++ b/src/everos/memory/strategies/reflect_episodes.py
@@ -22,7 +22,7 @@
 from everos.infra.ome.decorator import offline_strategy
 from everos.infra.ome.events import CronTick
 from everos.infra.ome.triggers import Cron
-from everos.infra.persistence.lancedb import (
+from everos.infra.persistence.index import (
     atomic_fact_repo,
     episode_repo,
 )
diff --git a/src/everos/service/get.py b/src/everos/service/get.py
index 9b9a7e679..3265f9eaa 100644
--- a/src/everos/service/get.py
+++ b/src/everos/service/get.py
@@ -1,8 +1,8 @@
 """Get use case — lazy singleton wiring for ``POST /api/v1/memory/get``.
 
 Mirrors :mod:`everos.service.search`: the :class:`GetManager` and its
-LanceDB repo singletons are built on first call so the FastAPI module
-import order stays decoupled from the lifespan that brings up LanceDB.
+derived index repo singletons are built on first call so the FastAPI module
+import order stays decoupled from the lifespan that brings up the index backend.
 
 ``/get`` is read-only and uses no embedding / LLM / rerank clients —
 it never blocks on optional components the way ``/search`` does.
@@ -11,7 +11,7 @@
 from __future__ import annotations
 
 from everos.core.observability.logging import get_logger
-from everos.infra.persistence.lancedb import (
+from everos.infra.persistence.index import (
     agent_case_repo,
     agent_skill_repo,
     episode_repo,
diff --git a/tests/unit/test_config/test_settings.py b/tests/unit/test_config/test_settings.py
index fb4fbb17f..b02d2e78b 100644
--- a/tests/unit/test_config/test_settings.py
+++ b/tests/unit/test_config/test_settings.py
@@ -120,6 +120,33 @@ def test_embedding_rerank_defaults() -> None:
     assert s.llm.api_key.get_secret_value() == ""
 
 
+def test_index_milvus_defaults() -> None:
+    s = Settings()
+    assert s.index.backend == "lancedb"
+    assert s.milvus.uri == ""
+    assert s.milvus.token.get_secret_value() == ""
+    assert s.milvus.db_name == ""
+    assert s.milvus.consistency_level == "Session"
+    assert s.milvus.dimension == 1024
+    assert s.milvus.collection_prefix == "everos"
+
+
+def test_index_milvus_env_overrides(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("EVEROS_INDEX__BACKEND", "milvus")
+    monkeypatch.setenv("EVEROS_MILVUS__URI", "http://localhost:19530")
+    monkeypatch.setenv("EVEROS_MILVUS__TOKEN", "secret")
+    monkeypatch.setenv("EVEROS_MILVUS__DB_NAME", "tenant_a")
+    monkeypatch.setenv("EVEROS_MILVUS__CONSISTENCY_LEVEL", "Strong")
+    monkeypatch.setenv("EVEROS_MILVUS__DIMENSION", "1536")
+    s = Settings()
+    assert s.index.backend == "milvus"
+    assert s.milvus.uri == "http://localhost:19530"
+    assert s.milvus.token.get_secret_value() == "secret"
+    assert s.milvus.db_name == "tenant_a"
+    assert s.milvus.consistency_level == "Strong"
+    assert s.milvus.dimension == 1536
+
+
 def test_resolve_root_default(monkeypatch: pytest.MonkeyPatch) -> None:
     """No --root, no EVEROS_ROOT → ~/.everos."""
     monkeypatch.delenv("EVEROS_ROOT", raising=False)
diff --git a/tests/unit/test_core/test_persistence/test_memory_root.py b/tests/unit/test_core/test_persistence/test_memory_root.py
index 8dd6dad15..cf36f24a3 100644
--- a/tests/unit/test_core/test_persistence/test_memory_root.py
+++ b/tests/unit/test_core/test_persistence/test_memory_root.py
@@ -67,6 +67,8 @@ def test_dotfile_paths(tmp_path: Path) -> None:
     mr = MemoryRoot(tmp_path)
     assert mr.index_dir == tmp_path / ".index"
     assert mr.lancedb_dir == tmp_path / ".index" / "lancedb"
+    assert mr.milvus_dir == tmp_path / ".index" / "milvus"
+    assert mr.milvus_db == tmp_path / ".index" / "milvus" / "milvus.db"
     assert mr.sqlite_dir == tmp_path / ".index" / "sqlite"
     assert mr.system_db == tmp_path / ".index" / "sqlite" / "system.db"
     assert mr.lock_file == tmp_path / ".lock"
@@ -80,6 +82,7 @@ def test_ensure_creates_required_dirs(tmp_path: Path) -> None:
     assert mr.index_dir.is_dir()
     assert mr.sqlite_dir.is_dir()
     assert mr.lancedb_dir.is_dir()
+    assert mr.milvus_dir.is_dir()
     assert mr.tmp_dir.is_dir()
     # User-visible dirs are NOT pre-created.
     assert not mr.agents_dir().exists()
diff --git a/tests/unit/test_infra/test_milvus/test_repo.py b/tests/unit/test_infra/test_milvus/test_repo.py
new file mode 100644
index 000000000..8d265d81f
--- /dev/null
+++ b/tests/unit/test_infra/test_milvus/test_repo.py
@@ -0,0 +1,128 @@
+"""Real Milvus Lite coverage for the optional derived index backend."""
+
+from __future__ import annotations
+
+import datetime as dt
+import importlib.util
+import uuid
+from pathlib import Path
+
+import pytest
+import pytest_asyncio
+
+from everos.config import load_settings
+
+pytestmark = pytest.mark.skipif(
+    importlib.util.find_spec("pymilvus") is None,
+    reason="pymilvus optional extra is not installed",
+)
+
+
+def test_milvus_score_normalization_handles_lite_and_cloud(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    from everos.infra.persistence.milvus import repository
+
+    monkeypatch.delenv("EVEROS_MILVUS__URI", raising=False)
+    monkeypatch.setattr(repository, "_milvus_lite_version", lambda: "3.0")
+    load_settings.cache_clear()
+    assert repository._cosine_distance_from_milvus(0.25) == pytest.approx(0.25)
+    assert repository._bm25_score_from_distance(-1.5) == pytest.approx(1.5)
+
+    monkeypatch.setattr(repository, "_milvus_lite_version", lambda: "3.0.0")
+    assert repository._cosine_distance_from_milvus(0.25) == pytest.approx(0.25)
+
+    monkeypatch.setenv("EVEROS_MILVUS__URI", "https://example.zillizcloud.com")
+    load_settings.cache_clear()
+    assert repository._cosine_distance_from_milvus(0.75) == pytest.approx(0.25)
+    assert repository._bm25_score_from_distance(1.5) == pytest.approx(1.5)
+
+    monkeypatch.delenv("EVEROS_MILVUS__URI", raising=False)
+    monkeypatch.setattr(repository, "_milvus_lite_version", lambda: "3.0.1")
+    load_settings.cache_clear()
+    assert repository._cosine_distance_from_milvus(0.75) == pytest.approx(0.25)
+
+
+@pytest_asyncio.fixture(autouse=True)
+async def _milvus_env(monkeypatch: pytest.MonkeyPatch, tmp_path: Path):
+    monkeypatch.setenv("EVEROS_ROOT", str(tmp_path))
+    monkeypatch.setenv("EVEROS_INDEX__BACKEND", "milvus")
+    monkeypatch.setenv("EVEROS_MILVUS__COLLECTION_PREFIX", f"test_{uuid.uuid4().hex}")
+    load_settings.cache_clear()
+    yield
+    from everos.infra.persistence.index import shutdown
+
+    await shutdown()
+    load_settings.cache_clear()
+
+
+async def test_milvus_episode_repo_upsert_search_filter_and_delete() -> None:
+    from everos.infra.persistence.index import Episode, episode_repo
+    from everos.memory.search.filters import compile_filters_for_backends
+
+    v_apple = [0.0] * 1024
+    v_apple[0] = 1.0
+    v_banana = [0.0] * 1024
+    v_banana[1] = 1.0
+    now = dt.datetime(2026, 1, 1, tzinfo=dt.UTC)
+
+    await episode_repo.upsert(
+        [
+            Episode(
+                id="u1_ep1",
+                entry_id="ep1",
+                owner_id="u1",
+                owner_type="user",
+                app_id="default",
+                project_id="default",
+                session_id="s1",
+                timestamp=now,
+                parent_id="mc1",
+                sender_ids=["user"],
+                episode="red apple memory",
+                episode_tokens="red apple memory",
+                md_path="default_app/default_project/users/u1/episodes/day.md",
+                content_sha256="a",
+                vector=v_apple,
+            ),
+            Episode(
+                id="u1_ep2",
+                entry_id="ep2",
+                owner_id="u1",
+                owner_type="user",
+                app_id="default",
+                project_id="default",
+                session_id="s1",
+                timestamp=now + dt.timedelta(seconds=1),
+                parent_id="mc2",
+                sender_ids=["assistant"],
+                episode="blue banana memory",
+                episode_tokens="blue banana memory",
+                md_path="default_app/default_project/users/u1/episodes/day.md",
+                content_sha256="b",
+                vector=v_banana,
+            ),
+        ]
+    )
+
+    filters = compile_filters_for_backends(None, owner_id="u1", owner_type="user")
+    rows = await episode_repo.find_where(filters, limit=10)
+    assert {row.id for row in rows} == {"u1_ep1", "u1_ep2"}
+    assert rows[0].timestamp.utcoffset() == dt.timedelta(0)
+    assert await episode_repo.count() == 2
+
+    sparse = await episode_repo.sparse_search(
+        ["apple"], filters, columns=Episode.BM25_FIELDS, limit=5
+    )
+    assert [row["id"] for row in sparse] == ["u1_ep1"]
+    assert sparse[0]["_score"] > 0
+
+    dense = await episode_repo.dense_search(v_apple, filters, limit=5)
+    assert dense[0]["id"] == "u1_ep1"
+    assert dense[0]["_distance"] == pytest.approx(0.0)
+
+    deleted = await episode_repo.delete_by_md_path(
+        "default_app/default_project/users/u1/episodes/day.md"
+    )
+    assert deleted == 2
+    assert await episode_repo.count() == 0
diff --git a/tests/unit/test_memory/test_get/test_manager.py b/tests/unit/test_memory/test_get/test_manager.py
index cb4e16714..93830a688 100644
--- a/tests/unit/test_memory/test_get/test_manager.py
+++ b/tests/unit/test_memory/test_get/test_manager.py
@@ -56,7 +56,7 @@ class _StubRepo:
 
     async def find_where_paginated(
         self,
-        where: str,
+        where: Any,
         *,
         sort_by: str,
         descending: bool = True,
@@ -65,7 +65,7 @@ async def find_where_paginated(
         max_fetch: int = 20000,
     ) -> tuple[list[Any], int]:
         self.last = _CallRecord(
-            where=where,
+            where=str(where),
             sort_by=sort_by,
             descending=descending,
             page=page,
diff --git a/tests/unit/test_memory/test_search/test_filters.py b/tests/unit/test_memory/test_search/test_filters.py
index 808c2a382..59e85dd2e 100644
--- a/tests/unit/test_memory/test_search/test_filters.py
+++ b/tests/unit/test_memory/test_search/test_filters.py
@@ -9,6 +9,7 @@
     FilterNode,
     compile_filters,
 )
+from everos.memory.search.filters import compile_filters_for_backends
 
 # ── Base injection ───────────────────────────────────────────────────────
 
@@ -266,3 +267,35 @@ def test_compile_filters_excludes_deprecated_by_for_user() -> None:
 def test_compile_filters_omits_deprecated_by_for_agent() -> None:
     result = compile_filters(None, owner_id="agent_1", owner_type="agent")
     assert "deprecated_by" not in result
+
+
+# ── Backend-specific rendering ─────────────────────────────────────────
+
+
+def test_compile_filters_for_backends_preserves_lancedb_default() -> None:
+    filters = compile_filters_for_backends(None, owner_id="u_a", owner_type="user")
+    assert str(filters) == filters.lancedb
+    assert "owner_id = 'u_a'" in filters.lancedb
+    assert 'owner_id == "u_a"' in filters.milvus
+    assert "deprecated_by IS NULL" in filters.lancedb
+    assert "deprecated_by is null" in filters.milvus
+
+
+def test_compile_filters_for_milvus_timestamp_and_array() -> None:
+    node = FilterNode.model_validate(
+        {"timestamp": {"gte": 1704067200000}, "sender_id": "u_jason"}
+    )
+    filters = compile_filters_for_backends(node, owner_id="u_a", owner_type="user")
+    assert "timestamp_ms >= 1704067200000" in filters.milvus
+    assert 'array_contains(sender_ids, "u_jason")' in filters.milvus
+    assert "TIMESTAMP '" in filters.lancedb
+    assert "array_has(sender_ids, 'u_jason')" in filters.lancedb
+
+
+def test_compile_filters_for_milvus_escapes_string_literals() -> None:
+    node = FilterNode.model_validate({"session_id": "ses's"})
+    filters = compile_filters_for_backends(node, owner_id="al'ice", owner_type="user")
+    assert 'owner_id == "al\'ice"' in filters.milvus
+    assert 'session_id == "ses\'s"' in filters.milvus
+    assert "owner_id = 'al''ice'" in filters.lancedb
+    assert "session_id = 'ses''s'" in filters.lancedb
diff --git a/tests/unit/test_memory/test_search/test_manager.py b/tests/unit/test_memory/test_search/test_manager.py
index ff8817be3..7e1d08fdc 100644
--- a/tests/unit/test_memory/test_search/test_manager.py
+++ b/tests/unit/test_memory/test_search/test_manager.py
@@ -101,25 +101,25 @@ def __init__(self, sparse: list[Candidate], dense: list[Candidate]) -> None:
         self.last_where: str | None = None
 
     async def sparse_recall(
-        self, query: str, where: str, *, limit: int
+        self, query: str, where: Any, *, limit: int
     ) -> list[Candidate]:
-        self.last_where = where
+        self.last_where = str(where)
         return list(self._sparse[:limit])
 
     async def dense_recall(
-        self, vector: Sequence[float], where: str, *, limit: int
+        self, vector: Sequence[float], where: Any, *, limit: int
     ) -> list[Candidate]:
-        self.last_where = where
+        self.last_where = str(where)
         return list(self._dense[:limit])
 
     async def fetch_by_parent_ids(
-        self, parent_ids: Sequence[str], where: str
+        self, parent_ids: Sequence[str], where: Any
     ) -> list[Candidate]:
         by_parent = {str(c.metadata.get("parent_id", "")): c for c in self._dense}
         return [by_parent[p] for p in parent_ids if p in by_parent]
 
     async def fetch_by_entry_ids(
-        self, entry_ids: Sequence[str], where: str
+        self, entry_ids: Sequence[str], where: Any
     ) -> list[Candidate]:
         by_entry = {str(c.metadata.get("entry_id", "")): c for c in self._dense}
         return [by_entry[e] for e in entry_ids if e in by_entry]
@@ -147,7 +147,7 @@ async def dense_recall(self, *_: Any, **__: Any) -> list[Candidate]:
     async def facts_for_episodes(
         self,
         ep_to_parents: Mapping[str, Sequence[str]],
-        where: str,
+        where: Any,
         *,
         per_episode: int,
         query_vector: Any = None,
@@ -198,7 +198,7 @@ async def dense_recall(self, *_: Any, **__: Any) -> list[Candidate]:
         return list(self._dense)
 
     async def fetch_by_case_ids(
-        self, case_ids: Sequence[str], where: str, *, limit: int
+        self, case_ids: Sequence[str], where: Any, *, limit: int
     ) -> list[Candidate]:
         return list(self._by_case)
 
diff --git a/tests/unit/test_memory/test_search/test_recall_atomic_fact.py b/tests/unit/test_memory/test_search/test_recall_atomic_fact.py
index f2e80b88e..c3c71d433 100644
--- a/tests/unit/test_memory/test_search/test_recall_atomic_fact.py
+++ b/tests/unit/test_memory/test_search/test_recall_atomic_fact.py
@@ -29,6 +29,7 @@
     atomic_fact_repo,
     lancedb_manager,
 )
+from everos.memory.search.recall import atomic_fact as atomic_fact_module
 from everos.memory.search.recall.atomic_fact import AtomicFactRecaller
 from everos.memory.search.recall.base import RecallerDeps
 
@@ -267,6 +268,34 @@ async def test_facts_for_episodes_score_zero_without_query_vector() -> None:
     assert out["alice_ep_a"][0].score == 0.0
 
 
+async def test_facts_for_episodes_caps_dense_recall_limit(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Large parent fan-out must not exceed Milvus/Zilliz search topK limits."""
+    captured: dict[str, int] = {}
+
+    async def fake_dense_search(vector, where, *, limit):  # type: ignore[no-untyped-def]
+        captured["limit"] = limit
+        return []
+
+    monkeypatch.setattr(
+        atomic_fact_module.atomic_fact_repo,
+        "dense_search",
+        fake_dense_search,
+    )
+    ep_to_parents = {f"alice_ep_{i}": [f"parent_{i}"] for i in range(80)}
+
+    out = await _recaller().facts_for_episodes(
+        ep_to_parents,
+        "owner_id = 'alice' AND owner_type = 'user'",
+        per_episode=20,
+        query_vector=_unit_vector(0),
+    )
+
+    assert out == {}
+    assert captured["limit"] == 1024
+
+
 # ── Dual parent_id (post-1.5 migration) ────────────────────────────────
 
 
diff --git a/tests/unit/test_memory/test_search/test_recall_episode.py b/tests/unit/test_memory/test_search/test_recall_episode.py
index 4549322d8..2cde9f368 100644
--- a/tests/unit/test_memory/test_search/test_recall_episode.py
+++ b/tests/unit/test_memory/test_search/test_recall_episode.py
@@ -58,9 +58,9 @@ async def test_fetch_all_for_owner_returns_entry_id_keyed_candidates(
         _make_row("ep_2", "mc_2"),
     ]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.search",
         new_callable=AsyncMock,
-        return_value=_mock_table(rows),
+        return_value=rows,
     ):
         result = await recaller.fetch_all_for_owner("owner_id = 'alice'")
 
@@ -75,9 +75,9 @@ async def test_fetch_all_for_owner_stores_episode_id_in_metadata(
     """metadata['episode_id'] carries the real LanceDB episode id for final shaping."""
     rows = [_make_row("ep_abc", "mc_xyz")]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.search",
         new_callable=AsyncMock,
-        return_value=_mock_table(rows),
+        return_value=rows,
     ):
         result = await recaller.fetch_all_for_owner("owner_id = 'alice'")
 
@@ -104,9 +104,9 @@ async def test_fetch_all_for_owner_skips_rows_without_entry_id(
         },
     ]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.search",
         new_callable=AsyncMock,
-        return_value=_mock_table(rows),
+        return_value=rows,
     ):
         result = await recaller.fetch_all_for_owner("owner_id = 'alice'")
 
@@ -131,9 +131,9 @@ async def test_fetch_all_for_owner_merged_episode_uses_entry_id(
         ),
     ]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.search",
         new_callable=AsyncMock,
-        return_value=_mock_table(rows),
+        return_value=rows,
     ):
         result = await recaller.fetch_all_for_owner("owner_id = 'alice'")
 
@@ -156,9 +156,9 @@ async def test_fetch_all_for_owner_mixed_regular_and_merged(
         ),
     ]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.search",
         new_callable=AsyncMock,
-        return_value=_mock_table(rows),
+        return_value=rows,
     ):
         result = await recaller.fetch_all_for_owner("owner_id = 'alice'")
 
@@ -179,14 +179,10 @@ async def test_fetch_by_entry_ids_returns_candidates(
             entry_id="entry_xyz",
         ),
     ]
-    mock_tbl = MagicMock()
-    mock_tbl.query.return_value.where.return_value.limit.return_value.to_list = (
-        AsyncMock(return_value=rows)
-    )
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.search",
         new_callable=AsyncMock,
-        return_value=mock_tbl,
+        return_value=rows,
     ):
         result = await recaller.fetch_by_entry_ids(["entry_xyz"], "owner_id = 'alice'")
 
@@ -226,9 +222,9 @@ async def test_sparse_recall_as_child_injects_parent_id(
         {**_make_row("ep_1", "mc_1", entry_id="entry_1"), "_score": 1.0},
     ]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.sparse_search",
         new_callable=AsyncMock,
-        return_value=_mock_bm25_table(rows),
+        return_value=rows,
     ):
         result = await recaller.sparse_recall_as_child(
             "hello", "owner_id = 'alice'", limit=10
@@ -256,9 +252,9 @@ async def test_sparse_recall_as_child_falls_back_to_id_when_no_entry_id(
         "_score": 0.5,
     }
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.sparse_search",
         new_callable=AsyncMock,
-        return_value=_mock_bm25_table([row]),
+        return_value=[row],
     ):
         result = await recaller.sparse_recall_as_child(
             "hello", "owner_id = 'alice'", limit=10
@@ -286,9 +282,9 @@ async def test_dense_recall_as_child_injects_parent_id(
         {**_make_row("ep_3", "mc_3", entry_id="entry_3"), "_distance": 0.1},
     ]
     with patch(
-        "everos.memory.search.recall.episode.get_table",
+        "everos.memory.search.recall.episode.episode_repo.dense_search",
         new_callable=AsyncMock,
-        return_value=_mock_ann_table(rows),
+        return_value=rows,
     ):
         result = await recaller.dense_recall_as_child(
             [0.1] * 1024, "owner_id = 'alice'", limit=10
@@ -325,10 +321,16 @@ async def test_dense_recall_subject_returns_subject_vector_source(
     rows = [
         {**_make_row("ep_s1", "mc_s1", entry_id="entry_s1"), "_distance": 0.2},
     ]
-    with patch(
-        "everos.memory.search.recall.episode.get_table",
-        new_callable=AsyncMock,
-        return_value=_mock_subject_ann_table(rows),
+    with (
+        patch(
+            "everos.memory.search.recall.episode.active_backend",
+            return_value="lancedb",
+        ),
+        patch(
+            "everos.infra.persistence.lancedb.get_table",
+            new_callable=AsyncMock,
+            return_value=_mock_subject_ann_table(rows),
+        ),
     ):
         result = await recaller.dense_recall_subject(
             [0.1] * 1024, "owner_id = 'alice'", limit=10
@@ -354,10 +356,16 @@ async def test_dense_recall_subject_as_child_injects_parent_id(
     rows = [
         {**_make_row("ep_s2", "mc_s2", entry_id="entry_s2"), "_distance": 0.15},
     ]
-    with patch(
-        "everos.memory.search.recall.episode.get_table",
-        new_callable=AsyncMock,
-        return_value=_mock_subject_ann_table(rows),
+    with (
+        patch(
+            "everos.memory.search.recall.episode.active_backend",
+            return_value="lancedb",
+        ),
+        patch(
+            "everos.infra.persistence.lancedb.get_table",
+            new_callable=AsyncMock,
+            return_value=_mock_subject_ann_table(rows),
+        ),
     ):
         result = await recaller.dense_recall_subject_as_child(
             [0.1] * 1024, "owner_id = 'alice'", limit=10
diff --git a/tests/unit/test_memory/test_search/test_recall_knowledge_topic.py b/tests/unit/test_memory/test_search/test_recall_knowledge_topic.py
index d40d8a41e..3e98c58b4 100644
--- a/tests/unit/test_memory/test_search/test_recall_knowledge_topic.py
+++ b/tests/unit/test_memory/test_search/test_recall_knowledge_topic.py
@@ -1,11 +1,11 @@
 """Unit tests for ``KnowledgeTopicRecaller``.
 
-Verifies dual-column BM25 + cosine ANN recall, using ``unittest.mock``
-to patch ``get_table`` so no real LanceDB connection is needed.
+Verifies dual-column BM25 delegation + cosine ANN recall, using
+``unittest.mock`` to patch the backend-neutral index repo.
 
 White-box surfaces touched:
-  - ``everos.memory.search.recall.knowledge_topic.get_table`` (patched)
-  - ``KnowledgeTopicRecaller.sparse_recall`` — queries both BM25 columns
+  - ``everos.memory.search.recall.knowledge_topic.knowledge_topic_repo`` (patched)
+  - ``KnowledgeTopicRecaller.sparse_recall`` — passes both BM25 columns
   - ``KnowledgeTopicRecaller.dense_recall`` — cosine ANN with distance→score
 """
 
@@ -17,6 +17,7 @@
 import pytest
 
 from everos.component.tokenizer import Tokenizer
+from everos.infra.persistence.index import KnowledgeTopic
 from everos.memory.search.recall.base import RecallerDeps
 from everos.memory.search.recall.knowledge_topic import KnowledgeTopicRecaller
 
@@ -58,41 +59,6 @@ def _make_row(
     return row
 
 
-def _mock_bm25_table(
-    summary_rows: list[dict[str, Any]],
-    content_rows: list[dict[str, Any]],
-) -> MagicMock:
-    """Build a table mock whose BM25 results differ per column.
-
-    The first ``nearest_to_text`` call (summary_tokens) returns
-    ``summary_rows``; the second (content_tokens) returns ``content_rows``.
-    ``asyncio.gather`` fires both concurrently, so we use ``side_effect``
-    on the chain rather than recording call order.
-    """
-    summary_chain = MagicMock()
-    summary_chain.where.return_value.limit.return_value.to_list = AsyncMock(
-        return_value=summary_rows
-    )
-
-    content_chain = MagicMock()
-    content_chain.where.return_value.limit.return_value.to_list = AsyncMock(
-        return_value=content_rows
-    )
-
-    tbl = MagicMock()
-    tbl.query.return_value.nearest_to_text.side_effect = [summary_chain, content_chain]
-    return tbl
-
-
-def _mock_ann_table(rows: list[dict[str, Any]]) -> MagicMock:
-    """Build a table mock for ANN (dense) queries."""
-    tbl = MagicMock()
-    ann = tbl.query.return_value.nearest_to.return_value
-    chain = ann.distance_type.return_value.where.return_value.limit.return_value
-    chain.to_list = AsyncMock(return_value=rows)
-    return tbl
-
-
 @pytest.fixture()
 def recaller() -> KnowledgeTopicRecaller:
     return KnowledgeTopicRecaller(RecallerDeps(tokenizer=_WhitespaceTokenizer()))
@@ -109,16 +75,20 @@ def recaller() -> KnowledgeTopicRecaller:
 async def test_sparse_recall_queries_both_columns(
     recaller: KnowledgeTopicRecaller,
 ) -> None:
-    """``nearest_to_text`` must be called once per BM25 column."""
-    tbl = _mock_bm25_table(
-        summary_rows=[_make_row("t1", score=0.9)],
-        content_rows=[_make_row("t2", score=0.7)],
-    )
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock, return_value=tbl):
+    """Both BM25 columns must be delegated to the index repo."""
+    rows = [_make_row("t1", score=0.9), _make_row("t2", score=0.7)]
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.sparse_search",
+        new_callable=AsyncMock,
+        return_value=rows,
+    ) as mock_sparse:
         result = await recaller.sparse_recall("topic query", _WHERE, limit=10)
 
-    # nearest_to_text called twice (once per column)
-    assert tbl.query.return_value.nearest_to_text.call_count == 2
+    mock_sparse.assert_awaited_once()
+    assert list(mock_sparse.await_args.args[0]) == ["topic", "query"]
+    assert mock_sparse.await_args.args[1] == _WHERE
+    assert mock_sparse.await_args.kwargs["columns"] == KnowledgeTopic.BM25_FIELDS
+    assert mock_sparse.await_args.kwargs["limit"] == 10
     ids = {c.id for c in result}
     assert ids == {"t1", "t2"}
 
@@ -126,13 +96,14 @@ async def test_sparse_recall_queries_both_columns(
 async def test_sparse_recall_merges_by_max_score(
     recaller: KnowledgeTopicRecaller,
 ) -> None:
-    """When the same id appears in both columns, keep the higher score."""
+    """Scores returned by the repo are preserved on keyword candidates."""
     shared_id = "topic_shared"
-    summary_rows = [_make_row(shared_id, score=0.5)]
-    content_rows = [_make_row(shared_id, score=0.9)]
 
-    tbl = _mock_bm25_table(summary_rows, content_rows)
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock, return_value=tbl):
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.sparse_search",
+        new_callable=AsyncMock,
+        return_value=[_make_row(shared_id, score=0.9)],
+    ):
         result = await recaller.sparse_recall("overlap", _WHERE, limit=10)
 
     assert len(result) == 1
@@ -144,16 +115,16 @@ async def test_sparse_recall_merges_by_max_score(
 async def test_sparse_recall_returns_sorted_by_score(
     recaller: KnowledgeTopicRecaller,
 ) -> None:
-    """Merged results must be sorted descending by score, truncated to limit."""
-    summary_rows = [
-        _make_row("a", score=0.3),
+    """The recaller preserves repo ordering and maps rows to candidates."""
+    rows = [
         _make_row("b", score=0.8),
-    ]
-    content_rows = [
         _make_row("c", score=0.6),
     ]
-    tbl = _mock_bm25_table(summary_rows, content_rows)
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock, return_value=tbl):
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.sparse_search",
+        new_callable=AsyncMock,
+        return_value=rows,
+    ):
         result = await recaller.sparse_recall("query", _WHERE, limit=2)
 
     assert len(result) == 2
@@ -164,16 +135,19 @@ async def test_sparse_recall_returns_sorted_by_score(
 async def test_sparse_recall_empty_query_returns_empty(
     recaller: KnowledgeTopicRecaller,
 ) -> None:
-    """Empty tokenisation short-circuits — no LanceDB query is issued."""
+    """Empty tokenisation short-circuits — no repo query is issued."""
     tok = MagicMock(spec=Tokenizer)
     tok.tokenize.return_value = []
     r = KnowledgeTopicRecaller(RecallerDeps(tokenizer=tok))
 
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock) as mock_gt:
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.sparse_search",
+        new_callable=AsyncMock,
+    ) as mock_sparse:
         result = await r.sparse_recall("", _WHERE, limit=10)
 
     assert result == []
-    mock_gt.assert_not_called()
+    mock_sparse.assert_not_called()
 
 
 # ---------------------------------------------------------------------------
@@ -189,8 +163,11 @@ async def test_dense_recall_cosine_conversion(
         _make_row("t1", distance=0.2),
         _make_row("t2", distance=0.5),
     ]
-    tbl = _mock_ann_table(rows)
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock, return_value=tbl):
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.dense_search",
+        new_callable=AsyncMock,
+        return_value=rows,
+    ):
         result = await recaller.dense_recall([0.1] * 1024, _WHERE, limit=10)
 
     assert len(result) == 2
@@ -203,12 +180,15 @@ async def test_dense_recall_cosine_conversion(
 async def test_dense_recall_empty_vector_returns_empty(
     recaller: KnowledgeTopicRecaller,
 ) -> None:
-    """Empty vector short-circuits — no LanceDB query is issued."""
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock) as mock_gt:
+    """Empty vector short-circuits — no repo query is issued."""
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.dense_search",
+        new_callable=AsyncMock,
+    ) as mock_dense:
         result = await recaller.dense_recall([], _WHERE, limit=10)
 
     assert result == []
-    mock_gt.assert_not_called()
+    mock_dense.assert_not_called()
 
 
 async def test_dense_recall_metadata_excludes_noise_columns(
@@ -218,8 +198,11 @@ async def test_dense_recall_metadata_excludes_noise_columns(
     row = _make_row("t1", distance=0.3)
     row["vector"] = [0.0] * 1024
 
-    tbl = _mock_ann_table([row])
-    with patch(f"{_MODULE}.get_table", new_callable=AsyncMock, return_value=tbl):
+    with patch(
+        f"{_MODULE}.knowledge_topic_repo.dense_search",
+        new_callable=AsyncMock,
+        return_value=[row],
+    ):
         result = await recaller.dense_recall([0.1] * 1024, _WHERE, limit=5)
 
     assert len(result) == 1
diff --git a/uv.lock b/uv.lock
index 65fa5d51f..10f4743e1 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1,6 +1,14 @@
 version = 1
 revision = 3
 requires-python = ">=3.12"
+resolution-markers = [
+    "python_full_version >= '3.14' and sys_platform == 'win32'",
+    "python_full_version >= '3.14' and sys_platform == 'emscripten'",
+    "python_full_version >= '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32'",
+    "python_full_version < '3.14' and sys_platform == 'win32'",
+    "python_full_version < '3.14' and sys_platform == 'emscripten'",
+    "python_full_version < '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32'",
+]
 
 [[package]]
 name = "aiosqlite"
@@ -99,6 +107,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/1a/39/47f9197bdd44df24d67ac8893641e16f386c984a0619ef2ee4c51fbbc019/beautifulsoup4-4.14.3-py3-none-any.whl", hash = "sha256:0918bfe44902e6ad8d57732ba310582e98da931428d231a5ecb9e7c703a735bb", size = 107721, upload-time = "2025-11-30T15:08:24.087Z" },
 ]
 
+[[package]]
+name = "cachetools"
+version = "7.1.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f4/8b/0d3945a13955303b81272f759a0331e54c5c793da455e6f5706b89d2639c/cachetools-7.1.4.tar.gz", hash = "sha256:437f55a4e0c1b01a4f3077cc470e6991d47430970e36fbcb77e2be0df4fc1cd6", size = 40085, upload-time = "2026-05-21T22:40:43.376Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8c/7b/1fc1c09cc0756cf25861a3be10565915953876da48bb228fb9a672b20a42/cachetools-7.1.4-py3-none-any.whl", hash = "sha256:323dc4127934744db5b54eb4924482d7edafbf9554e820d1531c2e08c0e4ef54", size = 16761, upload-time = "2026-05-21T22:40:41.845Z" },
+]
+
 [[package]]
 name = "cairocffi"
 version = "1.7.1"
@@ -592,6 +609,9 @@ dependencies = [
 ]
 
 [package.optional-dependencies]
+milvus = [
+    { name = "pymilvus", extra = ["milvus-lite"] },
+]
 multimodal = [
     { name = "everalgo-parser", extra = ["svg"] },
 ]
@@ -629,6 +649,7 @@ requires-dist = [
     { name = "prometheus-client", specifier = ">=0.20.0" },
     { name = "pydantic", specifier = ">=2.7.1" },
     { name = "pydantic-settings", specifier = ">=2.0.0" },
+    { name = "pymilvus", extras = ["milvus-lite"], marker = "extra == 'milvus'", specifier = ">=3.0.0" },
     { name = "python-multipart", specifier = ">=0.0.7" },
     { name = "pyyaml", specifier = ">=6.0" },
     { name = "sqlmodel", specifier = ">=0.0.22" },
@@ -639,7 +660,7 @@ requires-dist = [
     { name = "watchdog", specifier = ">=4.0.0" },
     { name = "watchfiles", specifier = ">=0.21.0" },
 ]
-provides-extras = ["multimodal"]
+provides-extras = ["multimodal", "milvus"]
 
 [package.metadata.requires-dev]
 dev = [
@@ -663,6 +684,23 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/c1/ea/53f2148663b321f21b5a606bd5f191517cf40b7072c0497d3c92c4a13b1e/executing-2.2.1-py2.py3-none-any.whl", hash = "sha256:760643d3452b4d777d295bb167ccc74c64a81df23fb5e08eff250c425a4b2017", size = 28317, upload-time = "2025-09-01T09:48:08.5Z" },
 ]
 
+[[package]]
+name = "faiss-cpu"
+version = "1.14.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "packaging" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/83/b0/48c083d01b7b68c463c1d56507147a9d733f791e1c469a77215a872a9fb5/faiss_cpu-1.14.3-cp310-abi3-macosx_14_0_arm64.whl", hash = "sha256:a9369863290a3f0e033757e4c10577b6ef7431f1cede394dabd0a137e4e2ed45", size = 4768290, upload-time = "2026-06-13T02:19:03.427Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/34/6b04ef5bae3eada6b5a9457d7875cce041c040d53c890815cbd1e9821c65/faiss_cpu-1.14.3-cp310-abi3-macosx_15_0_x86_64.whl", hash = "sha256:f9d0e84d909194f63f027bbd3c1e35e905e48c9345c2db6e6f24da09a6bc5906", size = 6925734, upload-time = "2026-06-13T02:19:05.232Z" },
+    { url = "https://files.pythonhosted.org/packages/7c/8a/b451af4b3c6dd18749ecfb58ccb68503b77e49c9aa4a89d950e9d521e058/faiss_cpu-1.14.3-cp310-abi3-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1d734cfa9ac90b6a5dfed3a27cb706d05f22824703dafc3969b4e2071877a31c", size = 9661210, upload-time = "2026-06-13T02:19:06.99Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/ed/57335bc18c9e18677587bec9bf070c675b29c8e683e13f4def0440731ca0/faiss_cpu-1.14.3-cp310-abi3-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8780b526c06e57aad90a8c4655dfba2ff1b3195bd600ff4499752fc45159c9fc", size = 18506292, upload-time = "2026-06-13T02:19:09.621Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/d3/c6ca8c44a63b909e78aa8a69e14501c79c89613e62261d58455ea603d710/faiss_cpu-1.14.3-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:b28ba083e8c02f2c9be03783402537fd3f00d27e68799c44bf88931500ee12ec", size = 11238800, upload-time = "2026-06-13T02:19:12.821Z" },
+    { url = "https://files.pythonhosted.org/packages/93/5f/b405692913a301251749cb175cb3f564ed257fdaa80a22c9a36444d0095d/faiss_cpu-1.14.3-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:cdcb90850cb4b7c27d270839b37bcc0dc8d1fb6a62d1e13053e51ac95061ca25", size = 19237637, upload-time = "2026-06-13T02:19:15.531Z" },
+]
+
 [[package]]
 name = "fastapi"
 version = "0.139.0"
@@ -824,6 +862,47 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/80/8f/774ce522de6a7e70fbeceeaeb6fbe502f5dfb8365728fb3bb4cb23463da8/grimp-3.14-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:a424ad14d5deb56721ac24ab939747f72ab3d378d42e7d1f038317d33b052b77", size = 2515157, upload-time = "2025-12-10T17:54:55.874Z" },
 ]
 
+[[package]]
+name = "grpcio"
+version = "1.82.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/90/bc/656b89387d6f4ed7e0686c7b64c2ae7e554a759aa58122c8e5fb99392c32/grpcio-1.82.1.tar.gz", hash = "sha256:707b24abd90fcb1e45bcc080577da1dbf9971d107490589b9539af8e1e77b4b5", size = 13187300, upload-time = "2026-07-08T12:36:16.588Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/dc/88/d1350bf3343a2ed87d801584e40609f6c6bd3087926eeca03de50348cf4a/grpcio-1.82.1-cp312-cp312-linux_armv7l.whl", hash = "sha256:c09bd5fa0d5b1fbd773ec349fe61441c3e4ebf168c229aa7538a820bdfad6a58", size = 6144689, upload-time = "2026-07-08T12:34:55.567Z" },
+    { url = "https://files.pythonhosted.org/packages/e6/33/71875cdecd27c24ac1385d4783a09853f01b84a825a36aec2a2bc7d0d080/grpcio-1.82.1-cp312-cp312-macosx_11_0_universal2.whl", hash = "sha256:1eae24810720734598e3e6a1a528d5de0f265fe3fc86575e9ecce424b9ec7379", size = 11952034, upload-time = "2026-07-08T12:34:58.128Z" },
+    { url = "https://files.pythonhosted.org/packages/82/b2/d9125df3d8a140dec12cc82c05b7deafedeababcff6496f28b2fd5634d10/grpcio-1.82.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:a6bd5daf5bde7b24d7ad2cbaf8bf9eac620d96222016bb5e7ddde930dec0673f", size = 6710772, upload-time = "2026-07-08T12:35:01.33Z" },
+    { url = "https://files.pythonhosted.org/packages/88/9b/69e2d1627398b964f34437dc476a5aff5a2cc8e7f247d26272b5674b5faf/grpcio-1.82.1-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:1ecfde669cb687ac020d31ff76debe5dc7a62213335f02262eb6625628da1c03", size = 7450677, upload-time = "2026-07-08T12:35:03.926Z" },
+    { url = "https://files.pythonhosted.org/packages/e4/e7/8f855ca29c294956122a2a73023655b9b02602d5111dad2b9b00e7631c68/grpcio-1.82.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:011c8badee95734dee8bf05ce3464756a0ac3ebb8d443afd20c0e2b5e4640ad9", size = 6886855, upload-time = "2026-07-08T12:35:06.174Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/f0/fa87e85f49925f44c479d07e58b051e69bcfef6b6d5fbc6749d140f6730a/grpcio-1.82.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:b85f4564926fb23114d239392bdcae200db1e6179629edd7d7ab0ab89c96a197", size = 7501323, upload-time = "2026-07-08T12:35:08.49Z" },
+    { url = "https://files.pythonhosted.org/packages/67/55/2e0b10ae1d3ef9dcc480b91dc2158f4931fc4675d3af0a2836e39b2a744f/grpcio-1.82.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:2c0c8270833395644c3fe6b6a806397955a2bc0538000a19a78b90c05a6c16e0", size = 8536899, upload-time = "2026-07-08T12:35:10.975Z" },
+    { url = "https://files.pythonhosted.org/packages/bd/95/a3d8b0431fa221efc51ee39d73595ede74ba82a43b7c4313192e580face2/grpcio-1.82.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:2ba199205ff46c7778290fe1673c91ac8e7e45678dd5c86e9e56fa33ec8788f6", size = 7913892, upload-time = "2026-07-08T12:35:13.944Z" },
+    { url = "https://files.pythonhosted.org/packages/b8/92/f2651ec704d9852a56faef394775038afba435b50ce82ab2404d119c3355/grpcio-1.82.1-cp312-cp312-win32.whl", hash = "sha256:06127691866e295c14e84a1fb86356dd962254f6abd0da4ca4b001eea9e89438", size = 4240985, upload-time = "2026-07-08T12:35:16.048Z" },
+    { url = "https://files.pythonhosted.org/packages/96/4f/a5fe8bf0d0a1b24855f370293075c931f27de4eb55f0f158786095bf3c11/grpcio-1.82.1-cp312-cp312-win_amd64.whl", hash = "sha256:1fa3223a3a2e1db74f4c2b255189eb7ea875dfba56e221d252ee3fc7b204778e", size = 5001580, upload-time = "2026-07-08T12:35:18.689Z" },
+    { url = "https://files.pythonhosted.org/packages/1b/3e/496992d08c0aaa11272eb6228dc8ab947da01fe835de243cd00521bce4c4/grpcio-1.82.1-cp313-cp313-linux_armv7l.whl", hash = "sha256:b454a2d97bfab7565683a02345f86bd182ab69fd7c2bdb7414171e7538f266b1", size = 6146068, upload-time = "2026-07-08T12:35:21.365Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/8f/f263d6f14fdba6b56cfadd91fd3e158a52682b72c6016d1f8723d435659f/grpcio-1.82.1-cp313-cp313-macosx_11_0_universal2.whl", hash = "sha256:3dde70abfc80b3be11de53ba0d601c439e7fb2afd3583ad1788d1146bec92fdc", size = 11948600, upload-time = "2026-07-08T12:35:24.312Z" },
+    { url = "https://files.pythonhosted.org/packages/8c/14/3a02e6ee49c2d85bc15eaae321e0e11ab3542cad3c5b2de121ecce0c4296/grpcio-1.82.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:f5523099c98c292ea1ae08e617249db760c56a78f8deae879027fe7d1ffbcbf6", size = 6714591, upload-time = "2026-07-08T12:35:27.027Z" },
+    { url = "https://files.pythonhosted.org/packages/69/80/58e3738696f48ab7645347b98d8a7f93d10e00e6218388fbfcd6c9310e3d/grpcio-1.82.1-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:5e5c4dc0a59b0f8490a6bdfd6fc8395b9d8ad8a8407c7d67ca7b5bba15c0877f", size = 7454995, upload-time = "2026-07-08T12:35:29.599Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/6c/2557c1a889363072fbf2285ecd0e8c44860d4dbd60f017a32537c5b863e2/grpcio-1.82.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:c40d94ba820329cc191981bc22fa6f6eed0799c6d921f3c6709521d59d4a2fd7", size = 6888621, upload-time = "2026-07-08T12:35:32.38Z" },
+    { url = "https://files.pythonhosted.org/packages/d2/66/907706ccaff1223f1e10fd5b37fc16faead43392fccb4e786e7e390ac141/grpcio-1.82.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:4c816180e31e273caaec6f8bd86a8392499d5bbb26f41da44e3dce48bde69095", size = 7505069, upload-time = "2026-07-08T12:35:35.072Z" },
+    { url = "https://files.pythonhosted.org/packages/b3/7c/ff97b0d0f635987ee5ec80dfedafa1aad629303745d48e8637d10eec5b80/grpcio-1.82.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:e31fd780b261830720cb70b0fd8f0aa51d49e75a66d7464ad2e31d4b765f2580", size = 8535384, upload-time = "2026-07-08T12:35:37.954Z" },
+    { url = "https://files.pythonhosted.org/packages/62/9e/a97fddd970a8d1588cade06eca20443761c1858b0ad6590a5c835aa18062/grpcio-1.82.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:9d76152d7c31d7210d4a106e5d8b64da5bba5d6abf11be30e2f7b0a0c59bbcbf", size = 7910707, upload-time = "2026-07-08T12:35:40.797Z" },
+    { url = "https://files.pythonhosted.org/packages/20/e4/eaba1517888af483a88d449eb7566f0f7f63446d46f339c5891798435875/grpcio-1.82.1-cp313-cp313-win32.whl", hash = "sha256:38e9dcb5258226fb3282630b31b16a968df52c8c6ad514af540646e0a4578f8a", size = 4240363, upload-time = "2026-07-08T12:35:43.298Z" },
+    { url = "https://files.pythonhosted.org/packages/b0/42/66a98d47732e35290bef722f6149fed3709cd4cf61166f6f53a12f417302/grpcio-1.82.1-cp313-cp313-win_amd64.whl", hash = "sha256:3dbfb52c36d9511ac2b8e6c94fdde837b393ae520cc321f52a333a2deedf5a90", size = 5000980, upload-time = "2026-07-08T12:35:46.262Z" },
+    { url = "https://files.pythonhosted.org/packages/b4/cb/cf9ae9e164c6e6dc8a494faa9771763df9da150eefe19671009624d1559f/grpcio-1.82.1-cp314-cp314-linux_armv7l.whl", hash = "sha256:35f990f7784c8fd2872644f07f96ebb4d9e48e145a190ab80d0280af91a1bfb2", size = 6146901, upload-time = "2026-07-08T12:35:49.261Z" },
+    { url = "https://files.pythonhosted.org/packages/3b/2a/eccf26dbcfb7f7cab8027c5490a16c8937c5aa7a2ec20a3eab2cf7a43165/grpcio-1.82.1-cp314-cp314-macosx_11_0_universal2.whl", hash = "sha256:46536a4a1f4434df3c851b9254ff6fc7df5705b273681a15ca277d5921c178a0", size = 11954756, upload-time = "2026-07-08T12:35:52.196Z" },
+    { url = "https://files.pythonhosted.org/packages/ad/75/3b3b4a3cc9f084b026af96e1d3e539b1af29ec7f41ed0dfff3cb99cc8626/grpcio-1.82.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d6650a7c1ebb7921c70e12a385439a8118efb99e669fa9ed31cf25db1843937c", size = 6723087, upload-time = "2026-07-08T12:35:54.973Z" },
+    { url = "https://files.pythonhosted.org/packages/9c/8b/b0f0c9b1400a99a4da4c09b114f101b192f8f11192e76f620b8962f5d90b/grpcio-1.82.1-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.whl", hash = "sha256:b8e110c66df5204c0506d6c8787b35d48b8b699ef5aa366d6c4d67325c67fe9a", size = 7454542, upload-time = "2026-07-08T12:35:57.586Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/bd/428e38868382aa193697a5aa53973f29c58e58ba4268aa0c86a2715ee58b/grpcio-1.82.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f853eae07235a51a27bb5d6a9a175a59ca55dc9b99edc6ce2f76f07332d333ae", size = 6889588, upload-time = "2026-07-08T12:36:00.012Z" },
+    { url = "https://files.pythonhosted.org/packages/49/ce/03e01d5e10259bf5c08ee50570cc94724e79c956f61fd2f09b341af0956c/grpcio-1.82.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:60b0f2c95337694fc094b77d9f60f50566c84b5677393e342eb98daeee242d98", size = 7514166, upload-time = "2026-07-08T12:36:02.693Z" },
+    { url = "https://files.pythonhosted.org/packages/ff/59/278b4b600329e2ba3849f3c1ea3c820b3a01b38a7ad184ba09595e8d2733/grpcio-1.82.1-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:b064fc444812bdaa9825d33c26f8d732d63ee6a5d78557c1faf92c98687fed27", size = 8536166, upload-time = "2026-07-08T12:36:05.349Z" },
+    { url = "https://files.pythonhosted.org/packages/44/27/7ccf2ef00f27a8e47a79d641c8ceaf7d3028c7a03d9a97b4c8a9a783c086/grpcio-1.82.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:7d7ede11d747b4e1bd05e3bc0260e155b65a88735a895a10f6521f19b889511e", size = 7912572, upload-time = "2026-07-08T12:36:08.393Z" },
+    { url = "https://files.pythonhosted.org/packages/0d/be/33742482d2753f2d3a1b7641664b6622262d44f2f3b609f13425dd86d36f/grpcio-1.82.1-cp314-cp314-win32.whl", hash = "sha256:3d21f19838dc255ecbb79321b15ae9b98fbddff4c3d4aedb0a81bdd7f4ab572a", size = 4321856, upload-time = "2026-07-08T12:36:10.899Z" },
+    { url = "https://files.pythonhosted.org/packages/cc/67/03329c847172c78ddeb1eb9be6b444fdbc12775a84c958b27e427e7b926d/grpcio-1.82.1-cp314-cp314-win_amd64.whl", hash = "sha256:e20f1edbb15f99e3128ec86433f9785fd5a451d8f115e74fe0056134f092a9d5", size = 5141114, upload-time = "2026-07-08T12:36:13.595Z" },
+]
+
 [[package]]
 name = "h11"
 version = "0.16.0"
@@ -1260,6 +1339,21 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/b3/38/89ba8ad64ae25be8de66a6d463314cf1eb366222074cfda9ee839c56a4b4/mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8", size = 9979, upload-time = "2022-08-14T12:40:09.779Z" },
 ]
 
+[[package]]
+name = "milvus-lite"
+version = "3.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "faiss-cpu" },
+    { name = "grpcio" },
+    { name = "numpy" },
+    { name = "pyarrow" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/a3/9a/d80d260e6fe1246818a8ef782c374ba9c6ca46ca3b987c14eabe914ef805/milvus_lite-3.0.tar.gz", hash = "sha256:2c35d0d046b1faae3402cde1fb73d65f51ee8c6aba65f54de1dda46f7bb18b9b", size = 589749, upload-time = "2026-05-13T07:14:05.827Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/59/42/dee08ae3bfc1731572f193d00b248c9370b0b9dff12becb0ffd8b2ee8d56/milvus_lite-3.0-py3-none-any.whl", hash = "sha256:d9a094eab84bdaa4253da3721482282c939da1cce6f4e1759f947e8d3e53406e", size = 230490, upload-time = "2026-05-13T07:14:00.816Z" },
+]
+
 [[package]]
 name = "nodeenv"
 version = "1.10.0"
@@ -1339,6 +1433,59 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/9d/1c/5d43735b2553baae2a5e899dcbcd0670a86930d993184d72ca909bf11c9b/openai-2.36.0-py3-none-any.whl", hash = "sha256:143f6194b548dbc2c921af1f1b03b9f14c85fed8a75b5b516f5bcc11a2a50c63", size = 1302361, upload-time = "2026-05-07T17:33:15.063Z" },
 ]
 
+[[package]]
+name = "orjson"
+version = "3.11.9"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/7e/0c/964746fcafbd16f8ff53219ad9f6b412b34f345c75f384ad434ceaadb538/orjson-3.11.9.tar.gz", hash = "sha256:4fef17e1f8722c11587a6ef18e35902450221da0028e65dbaaa543619e68e48f", size = 5599163, upload-time = "2026-05-06T15:11:08.309Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/16/6d/11867a3ffa3a3608d84a4de51ef4dd0896d6b5cc9132fbe1daf593e677bc/orjson-3.11.9-cp312-cp312-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:9ef6fe90aadef185c7b128859f40beb24720b4ecea95379fc9000931179c3a49", size = 228515, upload-time = "2026-05-06T15:09:57.265Z" },
+    { url = "https://files.pythonhosted.org/packages/24/75/05912954c8b288f34fcf5cd4b9b071cb4f6e77b9961e175e56ebb258089f/orjson-3.11.9-cp312-cp312-macosx_15_0_arm64.whl", hash = "sha256:e5c9b8f28e726e97d97696c826bc7bea5d71cecd63576dba92924a32c1961291", size = 128409, upload-time = "2026-05-06T15:09:59.063Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/86/1c3a47df3bc8191ea9ac51603bbb872a95167a364320c269f2557911f406/orjson-3.11.9-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:26a473dbb4162108b27901492546f83c76fdcea3d0eadff00ae7a07e18dcce09", size = 132106, upload-time = "2026-05-06T15:10:00.798Z" },
+    { url = "https://files.pythonhosted.org/packages/d7/cf/b33b5f3e695ae7d63feef9d915c37cc3b8f465493dcd4f8e0b4c697a2366/orjson-3.11.9-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:011382e2a60fda9d46f1cdee31068cfc52ffe952b587d683ec0463002802a0f4", size = 127864, upload-time = "2026-05-06T15:10:02.15Z" },
+    { url = "https://files.pythonhosted.org/packages/31/6a/6cf69385a58208024fcb8c014e2141b8ce838aba6492b589f8acfff97fab/orjson-3.11.9-cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c2d3dc759490128c5c1711a53eeaa8ee1d437fd0038ffd2b6008abf46db3f882", size = 135213, upload-time = "2026-05-06T15:10:03.515Z" },
+    { url = "https://files.pythonhosted.org/packages/e8/f8/0b1bd3e8f2efcdd376af5c8cfd79eaf13f018080c0089c80ebd724e3c7fb/orjson-3.11.9-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d8ea516b3726d190e1b4297e6f4e7a8650347ae053868a18163b4dd3641d1fff", size = 145994, upload-time = "2026-05-06T15:10:05.083Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/59/dab79f61044c529d2c81aecdc589b1f833a1c8dec11ba3b1c2498a02ca7e/orjson-3.11.9-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:380cdce7ba24989af81d0a7013d0aaec5d0e2a21734c0e2681b1bc4f141957fe", size = 132744, upload-time = "2026-05-06T15:10:06.853Z" },
+    { url = "https://files.pythonhosted.org/packages/0e/a4/82b7a2fe5d8a67a59ed831b24d59a3d46ea7d207b66e1602d376541d94a6/orjson-3.11.9-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:be4fa4f0af7fa18951f7ab3fc2148e223af211bf03f59e1c6034ec3f97f21d61", size = 134014, upload-time = "2026-05-06T15:10:08.213Z" },
+    { url = "https://files.pythonhosted.org/packages/50/c7/375e83a76851b73b2e39f3bcf0e5a19e2b89bad13e5bca97d0b293d27f24/orjson-3.11.9-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:a8f5f8bc7ce7d59f08d9f99fa510c06496164a24cb5f3d34537dbd9ca30132e2", size = 141509, upload-time = "2026-05-06T15:10:09.595Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/7c/49d5d82a3d3097f641f094f552131f1e2723b0b8cb0fa2874ab65ecfffa6/orjson-3.11.9-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:4d7fde5501b944f83b3e665e1b31343ff6e154b15560a16b7130ea1e594a4206", size = 415127, upload-time = "2026-05-06T15:10:11.049Z" },
+    { url = "https://files.pythonhosted.org/packages/3a/dc/7446c538590d55f455647e5f3c61fc33f7108714e7afcffa6a2a033f8350/orjson-3.11.9-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:cde1a448023ba7d5bb4c01c5afb48894380b5e4956e0627266526587ef4e535f", size = 148025, upload-time = "2026-05-06T15:10:12.842Z" },
+    { url = "https://files.pythonhosted.org/packages/df/e5/4d2d8af06f788329b4f78f8cc3679bb395392fcaa1e4d8d3c33e85308fa4/orjson-3.11.9-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:71e63adb0e1f1ed5d9e168f50a91ceb93ae6420731d222dc7da5c69409aa47aa", size = 136943, upload-time = "2026-05-06T15:10:14.405Z" },
+    { url = "https://files.pythonhosted.org/packages/06/69/850264ccf6d80f6b174620d30a87f65c9b1490aba33fe6b62798e618cad3/orjson-3.11.9-cp312-cp312-win32.whl", hash = "sha256:2d057a602cdd19a0ad680417527c45b6961a095081c0f46fe0e03e304aac6470", size = 131606, upload-time = "2026-05-06T15:10:15.791Z" },
+    { url = "https://files.pythonhosted.org/packages/b9/d5/973a43fc9c55e20f2051e9830997649f669be0cb3ca52192087c0143f118/orjson-3.11.9-cp312-cp312-win_amd64.whl", hash = "sha256:59e403b1cc5a676da8eaf31f6254801b7341b3e29efa85f92b48d272637e77be", size = 127101, upload-time = "2026-05-06T15:10:17.129Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/ae/495470f0e4a18f73fa10b7f6b84b464ec4cc5291c4e0c7c2a6c400bef006/orjson-3.11.9-cp312-cp312-win_arm64.whl", hash = "sha256:9af678d6488357948f1f84c6cd1c1d397c014e1ae2f98ae082a44eb48f602624", size = 126736, upload-time = "2026-05-06T15:10:18.645Z" },
+    { url = "https://files.pythonhosted.org/packages/32/33/93fcc25907235c344ae73122f8a4e01d2d393ef062b4af7d2e2487a32c37/orjson-3.11.9-cp313-cp313-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:4bab1b2d6141fe7b32ae71dac905666ece4f94936efbfb13d55bb7739a3a6021", size = 228458, upload-time = "2026-05-06T15:10:20.079Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/27/b1e6dadb3c080313c03fdd8067b85e6a0460c7d8d6a1c3984ef77b904e4d/orjson-3.11.9-cp313-cp313-macosx_15_0_arm64.whl", hash = "sha256:844417969855fc7a41be124aafe83dc424592a7f77cd4501900c67307122b92c", size = 128368, upload-time = "2026-05-06T15:10:21.549Z" },
+    { url = "https://files.pythonhosted.org/packages/21/0f/c9ede0bf052f6b4051e64a7d4fa91b725cccf8321a6a786e86eb03519f00/orjson-3.11.9-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ffe02797b5e9f3a9d8292ddcd289b474ad13e81ad83cd1891a240811f1d2cb81", size = 132070, upload-time = "2026-05-06T15:10:23.371Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/26/d398e28048dc18205bbe812f2c88cb9b40313db2470778e25964796458fe/orjson-3.11.9-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:0e4eed3b200023042814d2fc8a5d2e880f13b52e1ed2485e83da4f3962f7dc1a", size = 127892, upload-time = "2026-05-06T15:10:24.714Z" },
+    { url = "https://files.pythonhosted.org/packages/66/60/52b0054c4c700d5aa7fc5b7ca96917400d8f061307778578e67a10e25852/orjson-3.11.9-cp313-cp313-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:8aff7da9952a5ad1cef8e68017724d96c7b9a66e99e91d6252e1b133d67a7b10", size = 135217, upload-time = "2026-05-06T15:10:26.084Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/97/1e3dc2b2a28b7b2528f403d2fc1d79ec5f39af3bc143ab65d3ec26426385/orjson-3.11.9-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4d4e98d6f3b8afed8bc8cd9718ec0cdf46661826beefb53fe8eafb37f2bf0362", size = 145980, upload-time = "2026-05-06T15:10:28.062Z" },
+    { url = "https://files.pythonhosted.org/packages/fc/39/31fbfe7850f2de32dee7e7e5c09f26d403ab01e440ac96001c6b01ad3c99/orjson-3.11.9-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3a81d52442a7c99b3662333235b3adf96a1715864658b35bb797212be7bddb97", size = 132738, upload-time = "2026-05-06T15:10:29.727Z" },
+    { url = "https://files.pythonhosted.org/packages/a1/08/dca0082dd2a194acb93e5457e73455388e2e2ca464a2672449a9ddbb679d/orjson-3.11.9-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4e39364e726a8fff737309aff059ff67d8a8c8d5b677be7bb49a8b3e84b7e218", size = 134033, upload-time = "2026-05-06T15:10:31.152Z" },
+    { url = "https://files.pythonhosted.org/packages/11/d4/5bdb0626801230139987385554c5d4c42255218ac906525bf4347f22cd95/orjson-3.11.9-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:4fd66214623f1b17501df9f0543bef0b833979ab5b6ded1e1d123222866aa8c9", size = 141492, upload-time = "2026-05-06T15:10:32.641Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/88/a21fb53b3ede6703aede6dce4710ed4111e5b201cfa6bbff5e544f9d47d7/orjson-3.11.9-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:8ecc30f10465fa1e0ce13fd01d9e22c316e5053a719a8d915d4545a09a5ff677", size = 415087, upload-time = "2026-05-06T15:10:34.438Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/57/1b30daf70f0d8180e9a73cefbfbdd99e4bf19eb020466502b01fba7e0e50/orjson-3.11.9-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:97db4c94a7db398a5bd636273324f0b3fd58b350bbbac8bb380ceb825a9b40f4", size = 148031, upload-time = "2026-05-06T15:10:36.358Z" },
+    { url = "https://files.pythonhosted.org/packages/04/83/45fbb6d962e260807f99441db9613cee868ceda4baceda59b3720a563f97/orjson-3.11.9-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:9f78cf8fec5bd627f4082b8dfeac7871b43d7f3274904492a43dab39f18a19a0", size = 136915, upload-time = "2026-05-06T15:10:38.013Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/cc/2d10025f9056d376e4127ec05a5808b218d46f035fdc08178a5411b34250/orjson-3.11.9-cp313-cp313-win32.whl", hash = "sha256:d4087e5c0209a0a8efe4de3303c234b9c44d1174161dcd851e8eea07c7560b32", size = 131613, upload-time = "2026-05-06T15:10:39.569Z" },
+    { url = "https://files.pythonhosted.org/packages/67/bd/2775ff28bfe883b9aa1ff348300542eb2ef1ee18d8ae0e3a49846817a865/orjson-3.11.9-cp313-cp313-win_amd64.whl", hash = "sha256:051b102c93b4f634e89f3866b07b9a9a98915ada541f4ec30f177067b2694979", size = 127086, upload-time = "2026-05-06T15:10:41.262Z" },
+    { url = "https://files.pythonhosted.org/packages/91/2b/d26799e580939e32a7da9a39531bc9e58e15ca32ffaa6a8cb3e9bb0d22cd/orjson-3.11.9-cp313-cp313-win_arm64.whl", hash = "sha256:cce9127885941bd28f080cecf1f1d288336b7e0d812c345b08be88b572796254", size = 126696, upload-time = "2026-05-06T15:10:42.651Z" },
+    { url = "https://files.pythonhosted.org/packages/8e/eb/5da01e356015aee6ecfa1187ced87aef51364e306f5e695dd52719bf0e78/orjson-3.11.9-cp314-cp314-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:b6ef1979adc4bc243523f1a2ba91418030a8e29b0a99cbe7e0e2d6807d4dce6e", size = 228465, upload-time = "2026-05-06T15:10:44.097Z" },
+    { url = "https://files.pythonhosted.org/packages/64/62/3e0e0c14c957133bcd855395c62b55ed4e3b0af23ffea11b032cb1dcbdb1/orjson-3.11.9-cp314-cp314-macosx_15_0_arm64.whl", hash = "sha256:f36b7f32c7c0db4a719f1fc5824db4a9c6f8bd1a354debb91faf26ebf3a4c71e", size = 128364, upload-time = "2026-05-06T15:10:45.839Z" },
+    { url = "https://files.pythonhosted.org/packages/5a/5a/07d8aa117211a8ed7630bda80c8c0b14d04e0f8dcf99bcf49656e4a710eb/orjson-3.11.9-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:08f4d8ebb44925c794e535b2bebc507cebf32209df81de22ae285fb0d8d66de0", size = 132063, upload-time = "2026-05-06T15:10:47.267Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/ec/4acaf21483e18aa945be74a474c74b434f284b549f275a0a39b9f98956e9/orjson-3.11.9-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6cc7923789694fd58f001cbcac7e47abc13af4d560ebbfcf3b41a8b1a0748124", size = 122356, upload-time = "2026-05-06T15:10:48.765Z" },
+    { url = "https://files.pythonhosted.org/packages/13/d8/5f0555e7638801323b7a75850f92e7dfa891bc84fe27a1ba4449170d1200/orjson-3.11.9-cp314-cp314-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ea5c46eb2d3af39e806b986f4b09d5c2706a1f5afde3cbf7544ce6616127173c", size = 129592, upload-time = "2026-05-06T15:10:50.13Z" },
+    { url = "https://files.pythonhosted.org/packages/b6/30/ed9860412a3603ceb3c5955bfd72d28b9d0e7ba6ed81add14f83d7114236/orjson-3.11.9-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f5d89a2ed90731df3be64bab0aa44f78bff39fdc9d71c291f4a8023aa46425b7", size = 140491, upload-time = "2026-05-06T15:10:51.582Z" },
+    { url = "https://files.pythonhosted.org/packages/d0/17/adc514dea7ac7c505527febf884934b815d34f0c7b8693c1a8b39c5c4a57/orjson-3.11.9-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:25e4aed0312d292c09f61af25bba34e0b2c88546041472b09088c39a4d828af1", size = 127309, upload-time = "2026-05-06T15:10:53.329Z" },
+    { url = "https://files.pythonhosted.org/packages/76/3e/c0b690253f0b82d86e99949af13533363acfb5432ecb5d53dd5b3bce9c34/orjson-3.11.9-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:aaea64f3f467d22e70eeed68bdccb3bc4f83f650446c4a03c59f2cba28a108db", size = 134030, upload-time = "2026-05-06T15:10:54.988Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/7a/bc82a0bb25e9faaf92dc4d9ef002732efc09737706af83e346788641d4a7/orjson-3.11.9-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:a028425d1b440c5d92a6be1e1a020739dfe67ea87d96c6dbe828c1b30041728b", size = 141482, upload-time = "2026-05-06T15:10:56.663Z" },
+    { url = "https://files.pythonhosted.org/packages/01/55/e69188b939f77d5d32a9833745ace31ea5ccae3ab613a1ec185d3cd2c4fb/orjson-3.11.9-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:5b192c6cf397e4455b11523c5cf2b18ed084c1bbd61b6c0926344d2129481972", size = 415178, upload-time = "2026-05-06T15:10:58.446Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/1a/b8a5a7ac527e80b9cb11d51e3f6689b709279183264b9ec5c7bc680bb8b5/orjson-3.11.9-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:ea407d4ccf5891d667d045fecae97a7a1e5e87b3b97f97ae1803c2e741130be0", size = 148089, upload-time = "2026-05-06T15:11:00.441Z" },
+    { url = "https://files.pythonhosted.org/packages/97/4e/00503f64204bf859b37213a63927028f30fb6268cd8677fb0a5ad48155e1/orjson-3.11.9-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:5f63aaf97afd9f6dec5b1a68e1b8da12bfccb4cb9a9a65c3e0b6c847849e7586", size = 136921, upload-time = "2026-05-06T15:11:02.176Z" },
+    { url = "https://files.pythonhosted.org/packages/0d/ba/a23b82a0a8d0ed7bed4e5f5035aae751cad4ff6a1e8d2ecd14d8860f5929/orjson-3.11.9-cp314-cp314-win32.whl", hash = "sha256:e30ab17845bb9fa54ccf67fa4f9f5282652d54faa6d17452f47d0f369d038673", size = 131638, upload-time = "2026-05-06T15:11:03.696Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/c3/0c6798456bade745c75c452342dabacce5798196483e77e643be1f53877d/orjson-3.11.9-cp314-cp314-win_amd64.whl", hash = "sha256:32ef5f4283a3be81913947d19608eacb7c6608026851123790cd9cc8982af34b", size = 127078, upload-time = "2026-05-06T15:11:05.123Z" },
+    { url = "https://files.pythonhosted.org/packages/16/21/5a3f1e8913103b703a436a5664238e5b965ec392b555fe68943ea3691e6b/orjson-3.11.9-cp314-cp314-win_arm64.whl", hash = "sha256:eebdbdeef0094e4f5aefa20dcd4eb2368ab5e7a3b4edea27f1e7b2892e009cf9", size = 126687, upload-time = "2026-05-06T15:11:06.602Z" },
+]
+
 [[package]]
 name = "packaging"
 version = "26.2"
@@ -1348,6 +1495,58 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/df/b2/87e62e8c3e2f4b32e5fe99e0b86d576da1312593b39f47d8ceef365e95ed/packaging-26.2-py3-none-any.whl", hash = "sha256:5fc45236b9446107ff2415ce77c807cee2862cb6fac22b8a73826d0693b0980e", size = 100195, upload-time = "2026-04-24T20:15:22.081Z" },
 ]
 
+[[package]]
+name = "pandas"
+version = "3.0.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+    { name = "python-dateutil" },
+    { name = "tzdata", marker = "sys_platform == 'emscripten' or sys_platform == 'win32'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f8/87/4341c6252d1c47b08768c3d25ac487362bf403f0313ddae4a2a26c9b1b4c/pandas-3.0.3.tar.gz", hash = "sha256:696a4a00a2a2a35d4e5deb3fc946641b96c944f02230e4f76137fe35d806c4fc", size = 4651414, upload-time = "2026-05-11T18:54:29.21Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/24/f1/392f8c5bfc16f66a0d2d41561c01627c228fe7ed2a0d056ef11315042570/pandas-3.0.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fed2ff7fd9779120e388e285fc029bd5cf9490cdd2e4166a9ee22c0e49a9ab09", size = 10357846, upload-time = "2026-05-11T18:52:36.143Z" },
+    { url = "https://files.pythonhosted.org/packages/cf/3d/b16412745651e855f357e5e66930248688378853a6e2698a214e331fba1f/pandas-3.0.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b168fc218fd80a6cbdbdbc1a97ddc7889ed057d7eb45f50d866ceab5f39904c4", size = 9899550, upload-time = "2026-05-11T18:52:38.976Z" },
+    { url = "https://files.pythonhosted.org/packages/31/a8/fa2535168fffcedf67f4f6de28d2dd903a747ca7c8ea6989451aaeb3a92f/pandas-3.0.3-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0383c72c75cdcca61a9e116e611143902dbfd08bff356829c2f6d1cf40a9ca8c", size = 10412965, upload-time = "2026-05-11T18:52:41.915Z" },
+    { url = "https://files.pythonhosted.org/packages/65/b6/09b01cdbc15224e2850365192d17b7bdebb8bdbd8780ed221fcdf0d9a515/pandas-3.0.3-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6dc0b3fd2169c9157deed50b4d519553a3655c8c6a96027136d654592be973a9", size = 10894600, upload-time = "2026-05-11T18:52:45.02Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/a4/2eb28f2fccb4ced4a2c79ab2a5dee9ade1ebf44922ebad6fea158c9f95d4/pandas-3.0.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:7e65d5407dc0b394f509699650e4a2ec01c0514f21850f453fa60f3be79a5dbf", size = 11422824, upload-time = "2026-05-11T18:52:48.058Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/45/830bb57f533a4604b355e07edcb8ea18cf88b5f94e5fca92f27052d7c597/pandas-3.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f8894dc474d648fe7b6ff0ca9b0bd73950d19952bc1a6534540762c5d79d305c", size = 11950889, upload-time = "2026-05-11T18:52:50.905Z" },
+    { url = "https://files.pythonhosted.org/packages/b9/c5/fc1b368f303087d20e8c9bf3d6ceb186263cfac0ade735cd938538bea839/pandas-3.0.3-cp312-cp312-win_amd64.whl", hash = "sha256:c7be265b62cef88e253a941e4698604973736dcfe242fdb5198f0f7bc473cdcc", size = 9755463, upload-time = "2026-05-11T18:52:53.386Z" },
+    { url = "https://files.pythonhosted.org/packages/86/bd/fda8f9705b1b09c6ebe14bfc0fa0e4ec8584d54ea673628f157ff55131af/pandas-3.0.3-cp312-cp312-win_arm64.whl", hash = "sha256:557409bc4178e70ee8d9ddb494798e51ebf6ea59330f6be22c51bab2a7db6c49", size = 9066158, upload-time = "2026-05-11T18:52:56.038Z" },
+    { url = "https://files.pythonhosted.org/packages/c5/90/62d8302883c44308c477e222c3daf7c813a34c8e96985882fbd53d964352/pandas-3.0.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:67b3b64c11910cfa29f4e94a14d3bff9ee693b6fc76055e7cad549cee0aec5fa", size = 10331071, upload-time = "2026-05-11T18:52:58.838Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/ae/6a6493c783a101f165e4356953ba3c74d6f77f0042fa7d753da9dfbb640c/pandas-3.0.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:39436b377d56d2a2e52d0395bdbee171f01068e99af5250509aceeb929f765c7", size = 9875690, upload-time = "2026-05-11T18:53:01.431Z" },
+    { url = "https://files.pythonhosted.org/packages/62/7c/5df8e9f56c69a2769fbe9382a5ef8f2658c007e376434e1e2cbb57ad895f/pandas-3.0.3-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d4be06d68f9ddcfc645b87534911da79a8fbffc7573c80e0edcf42a5020624d8", size = 10381634, upload-time = "2026-05-11T18:53:04.393Z" },
+    { url = "https://files.pythonhosted.org/packages/99/68/1237369725aa617bb358263d535803e3053fdbc593513ec5ed9c9896b5b6/pandas-3.0.3-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a4eeb6830daf35a71cc09649bd823e2b542dac246cdee9614c6e4bd65028cd6a", size = 10891243, upload-time = "2026-05-11T18:53:07.643Z" },
+    { url = "https://files.pythonhosted.org/packages/25/93/77d108e8af7222b4a503ebde0e30215b1c2e4f8e53a526431890f22d5586/pandas-3.0.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:1928e07221f82db493cd4af1e23c1bfca524a19a4699887975bff68f49a72bfb", size = 11388659, upload-time = "2026-05-11T18:53:10.634Z" },
+    { url = "https://files.pythonhosted.org/packages/d0/bd/eff5b4399f332ac386c853f6cd2bd3fa2ca0061b9f36ecd9c4d7c4265649/pandas-3.0.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:51b1fe551acb77dac643c6fda86084d8d446c10fe64b06a9cc29c4cc8540e7f2", size = 11942880, upload-time = "2026-05-11T18:53:13.536Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/20/559ace4200982c3887d0b86bfd0d856a2143ef8ddab63cc07934951a964c/pandas-3.0.3-cp313-cp313-win_amd64.whl", hash = "sha256:a82d532a3351d435432cd913edbccaf8b8e01d4dd0e5ced5a8d2e8ecd94c7e44", size = 9757091, upload-time = "2026-05-11T18:53:16.306Z" },
+    { url = "https://files.pythonhosted.org/packages/3a/66/69055a09fe200f29f922a3eeec4804611900b95f52d932ece3393c3c0c19/pandas-3.0.3-cp313-cp313-win_arm64.whl", hash = "sha256:275c14e0fce14a2ec20eee474aecd305478ea3c1e6f6a9d8fe219a165542717e", size = 9057282, upload-time = "2026-05-11T18:53:18.768Z" },
+    { url = "https://files.pythonhosted.org/packages/57/0e/efe801b0e6811e8e650cd21b7f2608e30f08a7067e2bf6e8752b0d56ee3c/pandas-3.0.3-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:46997386d528eb40376ecd6b033cf4a8a1e5282580f68f43de875b78cba2199d", size = 10767016, upload-time = "2026-05-11T18:53:21.227Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/dc/eb55135a1d5f0f0519f28da1f609a206d2cad1f9c35c32d51e38dd7261ae/pandas-3.0.3-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:261e308dfb22448384b7580cf719d2f998fe2966c92893c3e77d14008af1f066", size = 10420210, upload-time = "2026-05-11T18:53:23.982Z" },
+    { url = "https://files.pythonhosted.org/packages/c6/3e/b1d5d955ce33ffecb407465a60bc32769d74fcf68224b7ae67ae11d4dea4/pandas-3.0.3-cp313-cp313t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dd1a5d1def6a46002e964510bdc67c368aa0951df5d1d9f8365336f5a1f490cd", size = 10336126, upload-time = "2026-05-11T18:53:26.731Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/76/a01261711ab60a22d71b862f0de20e4c504bf80457270ad8cb42110f6abc/pandas-3.0.3-cp313-cp313t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d72828c20c6d6e83e1e22a6a3b47b326b71664112fa9705dcbccfd7a39b62085", size = 10728051, upload-time = "2026-05-11T18:53:29.125Z" },
+    { url = "https://files.pythonhosted.org/packages/e9/21/ea191195e587b18cf682e97f433f81b2d0fbe341380e80a3e0d6e4403c8e/pandas-3.0.3-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:d26cbe1fcfc12e8fd900e2454163e466b2d3af84f7c75481df7683ffc073d870", size = 11350796, upload-time = "2026-05-11T18:53:32.056Z" },
+    { url = "https://files.pythonhosted.org/packages/64/69/f0eaaf54939f0e8c6768fd06be9af2cef9b36048b96dfb9e1b2c685a807e/pandas-3.0.3-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:3e91cec1879ada0624fc3dc9953c5cbd60208e59c0db28f540c5d6d47502422f", size = 11799741, upload-time = "2026-05-11T18:53:34.985Z" },
+    { url = "https://files.pythonhosted.org/packages/45/a4/865e0e510cae5fc2194de4db28be638952de942571ba9125934fd9c01d47/pandas-3.0.3-cp313-cp313t-win_amd64.whl", hash = "sha256:08d789b41f87e0905880e293cedf6197ce71fe67cc081358b1e148a491b9bd13", size = 10499958, upload-time = "2026-05-11T18:53:37.857Z" },
+    { url = "https://files.pythonhosted.org/packages/86/54/effdcc3c0ff7a08037889200e148ebe94c16c4f653be078c7b3675955df1/pandas-3.0.3-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:3650109c0f22879df8bd6179ab9ee3d7f1d1d4e7e0094a3f0032d9f51e2e64ac", size = 10336065, upload-time = "2026-05-11T18:53:41.099Z" },
+    { url = "https://files.pythonhosted.org/packages/68/10/bf2d6738d72748b961a3751ab89522d58c54efc36a8e1a12161216cd45cf/pandas-3.0.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:bab900348131a7db1f69a7309ef141fd5680f1487094193bcbbb61791573bf8f", size = 9926101, upload-time = "2026-05-11T18:53:43.515Z" },
+    { url = "https://files.pythonhosted.org/packages/ae/e9/e35cf11c8a136e757b956f5f0efdcaa50aecde85ea055f1898dfc68262f3/pandas-3.0.3-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ba7e08b9ac1d54569cd1e256e3668975ed624d6826f7b68df0342b012007bddb", size = 10457553, upload-time = "2026-05-11T18:53:46.394Z" },
+    { url = "https://files.pythonhosted.org/packages/58/3b/1cdec6772bdbaf7b25dab360c59f03cadf05492dd724c6540af905389b07/pandas-3.0.3-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9d71c63ae4ebdbf70209742096f1fc46a83a0613c99d4b23766cced9ff8cd62a", size = 10914065, upload-time = "2026-05-11T18:53:49.134Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/c2/1ef644445fcd72e3627bceec77e3560636f87ddce4ed841afe76b83b5bf9/pandas-3.0.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:e3a2ec42c98ffa2565a67e08e218d06d72576d758d90facb7c00805194d8f360", size = 11459188, upload-time = "2026-05-11T18:53:52.527Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/49/4d8d4f42cbc9c4adc7a1870f269c02cbd6cd40d059622c06fb298addcbad/pandas-3.0.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:335f62418ed562cfc3c49e9e196375c28b729dcef8543abf4f9438e381bf3c76", size = 11982966, upload-time = "2026-05-11T18:53:55.043Z" },
+    { url = "https://files.pythonhosted.org/packages/38/55/792619469bab9882d8bbd5865d45a72f6478762d04a9af4bf0d08c503e95/pandas-3.0.3-cp314-cp314-win_amd64.whl", hash = "sha256:3c20a521bbb85902f79f7270c80a59e1b5452d96d170c034f207181870f97ac5", size = 9876755, upload-time = "2026-05-11T18:53:58.067Z" },
+    { url = "https://files.pythonhosted.org/packages/2a/af/33c469653b0ba03b50c3a98192d4c07f0c75c66b263ceb097fce0ee97d31/pandas-3.0.3-cp314-cp314-win_arm64.whl", hash = "sha256:a2d2dff8a04f3917b55ab3910c32990f8ddf7eceba114947838cefa976a68977", size = 9198658, upload-time = "2026-05-11T18:54:00.733Z" },
+    { url = "https://files.pythonhosted.org/packages/a2/fa/b8c257bd76b8bd060c3a9151c1fca05e9b9c5e3af5d0f549c0356f6d143d/pandas-3.0.3-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:0d589105b3c14645af1738ff279b2995102d8f7a03b0a66dc8d95550eb513e04", size = 10787242, upload-time = "2026-05-11T18:54:03.564Z" },
+    { url = "https://files.pythonhosted.org/packages/54/eb/f19206ffb0bf1919002969aa448b4702c6594845156a6f8050674855aac3/pandas-3.0.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:13fc1e853d9e04743d11ba75a985ccbc2a317fe07d8af61e445a6fd24dacd6a6", size = 10436369, upload-time = "2026-05-11T18:54:06.311Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/24/c7c39fb4fe22b71a0c2d78bf0c585c600092d85f94f086d2b3b2f6ca27e2/pandas-3.0.3-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:819959dab7bbd0049c15623fbac4e29a191b9528160a61fb1032242d8ced2d9c", size = 10358306, upload-time = "2026-05-11T18:54:09.085Z" },
+    { url = "https://files.pythonhosted.org/packages/16/ec/dd2a9eb7fa1204df88c0864164e35b228ac581062ac612ba0a67fd812e4c/pandas-3.0.3-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:60ae316d3fd75d1858d450d0db0103ea2be3e7d4a95ec2f064f7e2ae63f7b028", size = 10758394, upload-time = "2026-05-11T18:54:11.956Z" },
+    { url = "https://files.pythonhosted.org/packages/95/6e/00c61ea8e85b4f6d8d35e11852a1a4998fc7fafc91c6a602d1cc9c972d64/pandas-3.0.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:bd3a518890b400d32f9023722dc9a9a5c969f00b415419a3c06c043f09bb5d7d", size = 11375717, upload-time = "2026-05-11T18:54:14.539Z" },
+    { url = "https://files.pythonhosted.org/packages/31/89/8fc1c268969fac43688d65fd92e67df24bd128d53cb4d2eee534cd307399/pandas-3.0.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:9c39be2d709d01fa972a0cabc522389fceca4f3969332ba25a7d6c5802cf976a", size = 11828897, upload-time = "2026-05-11T18:54:17.146Z" },
+    { url = "https://files.pythonhosted.org/packages/56/3b/e7d20dea247a3e6dc0bd8a6953854afbedc03951def4e7371e05e7263e25/pandas-3.0.3-cp314-cp314t-win_amd64.whl", hash = "sha256:4db8c527972a821cf5286b40ccc57642a39bc62e62022b42f99f8a67fca8c3a1", size = 10900855, upload-time = "2026-05-11T18:54:19.72Z" },
+    { url = "https://files.pythonhosted.org/packages/0f/54/68a0978d1ef8502b8492099beaa6e7a0c1b32e3b5d4f677f5810cb08711c/pandas-3.0.3-cp314-cp314t-win_arm64.whl", hash = "sha256:b2c95f8bfc1ee412bf482605d7bfd30c12d1d26bd59fdd91efeef1d4718decb1", size = 9466464, upload-time = "2026-05-11T18:54:22.754Z" },
+]
+
 [[package]]
 name = "parso"
 version = "0.8.7"
@@ -1505,6 +1704,21 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/84/03/0d3ce49e2505ae70cf43bc5bb3033955d2fc9f932163e84dc0779cc47f48/prompt_toolkit-3.0.52-py3-none-any.whl", hash = "sha256:9aac639a3bbd33284347de5ad8d68ecc044b91a762dc39b7c21095fcd6a19955", size = 391431, upload-time = "2025-08-27T15:23:59.498Z" },
 ]
 
+[[package]]
+name = "protobuf"
+version = "7.35.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/da/01/9ef0afd7999eb9badb3a768b4aedd78c86d4c65cfaf1958ab276199e76b4/protobuf-7.35.1.tar.gz", hash = "sha256:ce115a26fe0c39a2c29973d914d327e516a6455464489fe3cd1e51a1b354f81a", size = 458717, upload-time = "2026-06-11T21:55:40.257Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/10/03/8aeeb7458d22546bf64b5250ca1daeb5ff757d900e8e4a7476c6f0db843e/protobuf-7.35.1-cp310-abi3-macosx_10_9_universal2.whl", hash = "sha256:24f857477359a85c0c235261b8ba905fd51b2562f4a64ca1df5473f29850cbf6", size = 433226, upload-time = "2026-06-11T21:55:31.719Z" },
+    { url = "https://files.pythonhosted.org/packages/37/4b/dfb89eb0e652a1ff073c39a59fb5e3a83cfe9b57a2c83fa6d78270101767/protobuf-7.35.1-cp310-abi3-manylinux2014_aarch64.whl", hash = "sha256:11d6b0ec246892d85215b0a13ca6e0233cf5284b68f0ac02646427f4ff88a799", size = 328847, upload-time = "2026-06-11T21:55:34.035Z" },
+    { url = "https://files.pythonhosted.org/packages/0f/58/dc12f2cd484951524af6e3382c785869b9b3fb5e52ee95ae23add53ee8f9/protobuf-7.35.1-cp310-abi3-manylinux2014_s390x.whl", hash = "sha256:b73f9489a4b8b1c9cb1f8ed951c736392592edb24b9d6819f36d2e10b171d5b4", size = 344030, upload-time = "2026-06-11T21:55:34.941Z" },
+    { url = "https://files.pythonhosted.org/packages/e4/be/5b3cfe508bfab6761414ff944e3366eb13be4fd71efcd69450f89ba39f43/protobuf-7.35.1-cp310-abi3-manylinux2014_x86_64.whl", hash = "sha256:74758715c53d7158fb76caf4f0cfdacc5329a4b1bb994f865d6cf302d413a1c4", size = 327130, upload-time = "2026-06-11T21:55:35.921Z" },
+    { url = "https://files.pythonhosted.org/packages/d8/bc/6d6c7ba8709c85f8f2c390b2b118d6fb08a783676a572271851bf45a7d22/protobuf-7.35.1-cp310-abi3-win32.whl", hash = "sha256:353652e4efd0bca5b5fc2656abf8307ef351f0cf938c9eba09f0e09c20a25c30", size = 428945, upload-time = "2026-06-11T21:55:37.034Z" },
+    { url = "https://files.pythonhosted.org/packages/0a/19/8d0cb6f20a1ef7b18f1c8986ad5783f22f84cce39c6ce9a6e645ea55192e/protobuf-7.35.1-cp310-abi3-win_amd64.whl", hash = "sha256:230a75ddfc2de4806e56696ce9640c1cdfdb6543b7cfce98d42a4c0a0e7bdb87", size = 439996, upload-time = "2026-06-11T21:55:38.123Z" },
+    { url = "https://files.pythonhosted.org/packages/19/c7/5f7c636ec43e0c545e28d1f1db71990108306f7bdcb89f069ba97e428e7f/protobuf-7.35.1-py3-none-any.whl", hash = "sha256:4bc97768d8fe4ad6743c8a19403e314511ed9f6d13205b687e52421c023ac1b9", size = 171659, upload-time = "2026-06-11T21:55:39.155Z" },
+]
+
 [[package]]
 name = "psutil"
 version = "7.2.2"
@@ -1756,6 +1970,29 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/8e/97/03635143a12a5d941f545548b00f8ac39d35565321a2effb4154ed267338/pyinstrument-5.1.2-cp314-cp314t-win_amd64.whl", hash = "sha256:b6a71f5e7f53c86c9b476b30cf19509463a63581ef17ddbd8680fee37ae509db", size = 128164, upload-time = "2026-01-04T18:38:32.281Z" },
 ]
 
+[[package]]
+name = "pymilvus"
+version = "3.0.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cachetools" },
+    { name = "grpcio" },
+    { name = "orjson" },
+    { name = "pandas" },
+    { name = "protobuf" },
+    { name = "python-dotenv" },
+    { name = "requests" },
+    { name = "setuptools" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ac/c1/01647e61f3a82fd881382746b6dde3401d65b88cd4f75bd059901fb2392b/pymilvus-3.0.0-1-py3-none-any.whl", hash = "sha256:57c8e7c87fbbf579f122b4df893949bc78e50bca2988527864891bd544817b05", size = 344817, upload-time = "2026-05-07T14:57:45.235Z" },
+]
+
+[package.optional-dependencies]
+milvus-lite = [
+    { name = "milvus-lite", marker = "sys_platform != 'win32'" },
+]
+
 [[package]]
 name = "pytest"
 version = "9.0.3"
@@ -2061,6 +2298,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/c0/98/6beb4b351e472e5f4c4613f7c35a5290b8be2497e183825310c4c3a3984b/ruff-0.15.12-py3-none-win_arm64.whl", hash = "sha256:a538f7a82d061cee7be55542aca1d86d1393d55d81d4fcc314370f4340930d4f", size = 11120821, upload-time = "2026-04-24T18:16:57.979Z" },
 ]
 
+[[package]]
+name = "setuptools"
+version = "83.0.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/34/26/f5d29e25ffdb535afef2d35cdb55b325298f96debd670da4c325e08d70f4/setuptools-83.0.0.tar.gz", hash = "sha256:025bccbbf0fa05b6192bc64ae1e7b16e001fd6d6d4d5de03c97b1c1ade523bef", size = 1154254, upload-time = "2026-07-04T15:31:22.699Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/40/e1e72872c6354b306daef1703549e8e83b4d43cfea356311bf722a043752/setuptools-83.0.0-py3-none-any.whl", hash = "sha256:29b23c360f22f414dc7336bb39178cc7bcbf6021ed2733cde173f09dba19abb3", size = 1008090, upload-time = "2026-07-04T15:31:20.885Z" },
+]
+
 [[package]]
 name = "shellingham"
 version = "1.5.4"