feat: local embedding provider for search

lat.md/cli.md

@@ -303,7 +303,9 @@ Core search logic in [[src/cli/search.ts#runSearch]] (returns matched sections),
 
 ### Provider Detection
 
-Requires an LLM key resolved by [[src/config.ts#getLlmKey]] in priority order:
+Resolves an embedding provider from the LLM key (or lack thereof). When no key is configured, falls back to a local embedding model so search works out of the box without any API key.
+
+Key resolution order (via [[src/config.ts#getLlmKey]]):
 
 1. `LAT_LLM_KEY` env var — direct value
 2. `LAT_LLM_KEY_FILE` env var — path to a file containing the key (read and trimmed)
@@ -312,14 +314,29 @@ Requires an LLM key resolved by [[src/config.ts#getLlmKey]] in priority order:
 
 Provider is auto-detected from the resolved key prefix:
 
+- *(no key)* — local embeddings via `@huggingface/transformers` (see [[cli#search#Local Embeddings]])
 - `sk-...` — OpenAI (uses `text-embedding-3-small`, 1536 dims)
 - `vck_...` — Vercel AI Gateway (uses `openai/text-embedding-3-small`, 1536 dims)
 - `sk-ant-...` — Anthropic (not supported, errors with guidance)
 - `REPLAY_LAT_LLM_KEY::<url>` — test-only replay server for offline testing
 
 Implementation: [[src/search/provider.ts]], [[src/config.ts]]
 
-### Embeddings
+### Local Embeddings
+
+Falls back to a local model when no API key is configured. Uses `@huggingface/transformers` (optional dep).
+
+Three pre-defined model sizes, selected via `LAT_LOCAL_MODEL_SIZE` env var (default `small`):
+
+- `small` — `Xenova/all-MiniLM-L6-v2`, 384 dims, ~45 MB download
+- `medium` — `Xenova/bge-base-en-v1.5`, 768 dims, ~130 MB download
+- `large` — `Xenova/bge-large-en-v1.5`, 1024 dims, ~330 MB download
+
+Dimensions are known statically from the model table — no need to load the model to discover them. The pipeline promise is cached so concurrent callers share a single load. Errors (missing package, invalid size) surface before any indexing work begins.
+
+Implementation: [[src/search/local.ts#embedLocal]]
+
+### API Embeddings
 
 Direct `fetch()` calls to the provider's OpenAI-compatible `/v1/embeddings` endpoint. No LangChain or other framework — keeps the dependency tree minimal. Batches up to 2048 texts per request.
 
@@ -331,6 +348,8 @@ Uses `@libsql/client` (Turso's libsql) in local file mode — pure JS/WASM, no n
 
 Single `sections` table holds metadata, content, content hash, and the embedding vector. No separate vector table needed.
 
+A `meta` table tracks the current embedding dimensions. On dimension mismatch (e.g. switching from API at 1536 to local at 384), the table is dropped and rebuilt in a single `db.batch()` call.
+
 The database is stored at `lat.md/.cache/vectors.db` and should not be committed (included in `.gitignore` template).
 
 Implementation: [[src/search/db.ts]]

lat.md/tests/mcp.md

@@ -40,5 +40,5 @@ Semantic search via `lat_search` for a login/security query returns results cont
 ## lat_search finds performance section
 Semantic search for a latency/response-times query returns results containing the Performance section.
 
-## lat_search returns no results message
-When `LAT_LLM_KEY` is not set, `lat_search` returns an error with `isError: true` explaining the missing key.
+## lat_search works without an API key
+When `LAT_LLM_KEY` is not set, `lat_search` handles the no-key case gracefully. If `@huggingface/transformers` is installed, returns search results via local embeddings. If not, returns a clean error with install guidance. Either way, no crash.

lat.md/tests/search.md

@@ -8,7 +8,19 @@ Tests in `tests/search.test.ts`.
 
 ## Provider Detection
 
-Unit tests (always run). Verify `detectProvider` correctly identifies OpenAI (`sk-`), Vercel (`vck_`), rejects Anthropic (`sk-ant-`) with a helpful message, and rejects unknown prefixes.
+Unit tests (always run). Verify `detectProvider` identifies providers by key prefix and returns local with correct dimensions when no key is given.
+
+Covers OpenAI (`sk-`), Vercel (`vck_`), local (no key, 384 dims), Anthropic rejection (`sk-ant-`), and unknown prefix rejection.
+
+## Schema Dimension Mismatch
+
+Verify that `ensureSchema` rebuilds the sections table when stored dimensions differ from the provider's.
+
+Creates a DB at 1536 dimensions with a row, re-inits at 384, asserts the table is empty and the new dimension is stored in meta.
+
+## Local Embedding
+
+Gated on `@huggingface/transformers` availability. Verifies that `embedLocal` produces normalized vectors with correct dimensions (384 for `Xenova/all-MiniLM-L6-v2`) and that semantically similar texts rank closer than unrelated ones.
 
 ## RAG Replay Tests
 

package.json

@@ -46,6 +46,9 @@
     "typescript": "^5.7.0",
     "vitest": "^3.0.0"
   },
+  "optionalDependencies": {
+    "@huggingface/transformers": "^3.8.1"
+  },
   "dependencies": {
     "@folder/xdg": "^4.0.1",
     "@libsql/client": "^0.17.0",