diff --git a/docs/site/.vitepress/config.mts b/docs/site/.vitepress/config.mts
index 27fe86c..eaebef0 100644
--- a/docs/site/.vitepress/config.mts
+++ b/docs/site/.vitepress/config.mts
@@ -88,11 +88,19 @@ export default defineConfig({
           ],
           '/ecosystem/': [
             {
-              text: 'Ecosystem',
+              text: 'Systems',
               items: [
                 { text: 'RAG Systems', link: '/ecosystem/rag-systems' },
                 { text: 'LLM Wiki Systems', link: '/ecosystem/llm-wiki-systems' },
               ]
+            },
+            {
+              text: 'Technologies',
+              items: [
+                { text: 'RAG Technologies', link: '/ecosystem/rag-technologies' },
+                { text: 'LLM Wiki Technologies', link: '/ecosystem/llm-wiki-technologies' },
+                { text: 'Paradigm Comparison', link: '/ecosystem/paradigm-comparison' },
+              ]
             }
           ],
         }
@@ -159,11 +167,19 @@ export default defineConfig({
           ],
           '/zh-CN/ecosystem/': [
             {
-              text: '生态',
+              text: '系统列表',
               items: [
                 { text: 'RAG 系统', link: '/zh-CN/ecosystem/rag-systems' },
                 { text: 'LLM Wiki 系统', link: '/zh-CN/ecosystem/llm-wiki-systems' },
               ]
+            },
+            {
+              text: '关键技术',
+              items: [
+                { text: 'RAG 关键技术', link: '/zh-CN/ecosystem/rag-technologies' },
+                { text: 'LLM Wiki 关键技术', link: '/zh-CN/ecosystem/llm-wiki-technologies' },
+                { text: '范式对比', link: '/zh-CN/ecosystem/paradigm-comparison' },
+              ]
             }
           ],
         }
diff --git a/docs/site/ecosystem/llm-wiki-technologies.md b/docs/site/ecosystem/llm-wiki-technologies.md
new file mode 100644
index 0000000..dc20f42
--- /dev/null
+++ b/docs/site/ecosystem/llm-wiki-technologies.md
@@ -0,0 +1,77 @@
+# LLM Wiki Technologies
+
+Core technologies behind the LLM Wiki (Knowledge Compilation) paradigm — knowledge is pre-compiled at write time, not retrieved at query time.
+
+```
+Raw docs → LLM distills → wiki pages → [query time] → FTS / graph → read compiled page
+```
+
+> Coined by Andrej Karpathy: use LLM as a compiler that transforms raw materials into a persistent, structured, auditable knowledge layer.
+
+## 1. Knowledge Compilation
+
+| Technology | Notes |
+|---|---|
+| LLM as compiler | LLM reads raw docs and writes structured wiki pages — pre-compiled, not retrieved at query time. |
+| Distillation pipeline | Raw doc → LLM extracts key claims, entities, aliases, related links → source-note page (WikiLoop pattern). |
+| Tiered compilation | Compile frequently-used docs first (Tier 0–3). Sage Wiki: auto-promote on 3 hits, auto-demote after 90 days inactive. |
+| Incremental update | Only re-process changed documents — avoids full rebuild cost. |
+| [STORM](https://github.com/stanford-oval/storm) (Stanford OVAL) | Multi-perspective Wiki generation agent: simulates different viewpoints → retrieves → generates hierarchical outline → writes cited article. |
+| Co-STORM | Collaborative variant of STORM — builds dynamic knowledge map during research to guide compilation direction. |
+| Agentic compilation loop | Agent loop: retrieve → draft → evaluate → re-retrieve → regenerate until knowledge is stable. |
+
+## 2. Knowledge Representation
+
+| Technology | Notes |
+|---|---|
+| Structured Markdown | All knowledge stored as plain Markdown — human-readable, git-diffable, auditable. |
+| Source-note pages | One distilled page per raw document. Contains `key_claims`, entity annotations `【entity\|type】`, `related_to` / `supports` / `contradicts` links. |
+| Concept pages | Cross-document synthesis of a concept or methodology. |
+| Comparison pages | Side-by-side comparison of approaches, tools, or paradigms. |
+| Decision pages | Technical decision records (ADR format): context + options + decision + consequences. |
+| Ontology graph | Typed entity-relation graph built during compilation. Sage Wiki: 8 built-in relation types (`implements`, `contradicts`, `trades_off`, …). |
+| Schema / Templates | Authoring rules and page templates that guide LLM compilation style, customizable per KB. |
+
+## 3. Indexing & Search
+
+| Technology | Notes |
+|---|---|
+| SQLite FTS5 + BM25 | Core search engine — no vector model needed. Sub-millisecond full-text search. Used by WikiLoop, Sage Wiki, TreeSearch. |
+| Alias expansion | Key terms indexed with aliases and cross-language equivalents to maximize FTS recall. |
+| Graph traversal | `related_to` links enable multi-hop navigation (similar to wiki page links). BFS expansion on search results. |
+| Hybrid (FTS + vector + graph) | Sage Wiki: FTS5 (411µs) + vector (81ms) + ontology graph (1µs) merged via RRF. |
+| Chapter-tree indexing | Preserves document H1/H2/H3 hierarchy — structure-preserving alternative to flat chunking. |
+
+## 4. Knowledge Quality & Maintenance
+
+| Technology | Notes |
+|---|---|
+| Git version control | All wiki pages in git — full history, diffs, blame. Knowledge changes are auditable. |
+| Lint / health checks | Validate frontmatter, broken source links, missing citations. `wikiloop lint`. |
+| Conflict detection | `contradicts` links surface disagreements between sources for human review. |
+| Draft staging | Pages with < 2 source references quarantined in `_draft/` — not indexed until verified. |
+| Knowledge gap analysis | `wikiloop synthesize --gaps` identifies topics with insufficient coverage. |
+| Entity deduplication | Identify different expressions of the same concept and merge into a single node. |
+
+## 5. Agent Interface (MCP)
+
+| Technology | Notes |
+|---|---|
+| [MCP protocol](https://modelcontextprotocol.io) | Model Context Protocol — Anthropic open standard for exposing tools/resources to AI agents. Supports stdio + HTTP transports. |
+| `kb_search` | FTS keyword search, returns ranked results with `related` links for graph navigation. |
+| `kb_page` | Fetch full page content by ID. Supports batch (up to 5 IDs) or `full=true` for untruncated text. |
+| MCP Resources | Read-only resources (URI form): wiki pages, graph schema, raw documents. |
+| Iterative search pattern | Agent issues multiple queries from different angles, follows `related` links, synthesizes own answer. |
+| Sage Wiki MCP tools | 17 tools: 6 read, 9 write, 2 composite — agents can directly write and compile knowledge. |
+
+## 6. File Conversion (Input Layer)
+
+| Technology | Notes |
+|---|---|
+| markitdown (Microsoft) | Converts PDF, Word, Excel, PPT, HTML to Markdown before distillation. |
+| `raw/converted/` pattern | Agent-extracted content written directly here — skips conversion, goes straight to distillation. |
+| File watcher | Auto-detects new/changed files in `raw/`, triggers convert → distill → index pipeline. |
+
+---
+
+See also: [RAG Technologies](/ecosystem/rag-technologies) · [Paradigm Comparison](/ecosystem/paradigm-comparison)
diff --git a/docs/site/ecosystem/paradigm-comparison.md b/docs/site/ecosystem/paradigm-comparison.md
new file mode 100644
index 0000000..d0b9baa
--- /dev/null
+++ b/docs/site/ecosystem/paradigm-comparison.md
@@ -0,0 +1,65 @@
+# Paradigm Comparison: RAG vs LLM Wiki
+
+Two fundamentally different approaches to building AI knowledge infrastructure.
+
+## Pipeline Overview
+
+```
+RAG Pipeline
+────────────────────────────────────────────────────────
+Raw docs → chunk → embed → vector store
+                                         ↘ [query time]
+                              query → ANN search → LLM → answer
+
+
+LLM Wiki (Knowledge Compilation)
+────────────────────────────────────────────────────────
+Raw docs → LLM compiles → structured wiki pages → FTS index
+                                                           ↘ [query time]
+                                    query → FTS / graph → read page → answer
+```
+
+## Dimension-by-Dimension Comparison
+
+| Dimension | RAG | LLM Wiki |
+|---|---|---|
+| **Core operation** | Retrieve at query time | Compile at write time |
+| **Storage** | Vector DB + raw chunks | Structured Markdown + SQLite FTS |
+| **Search** | ANN similarity search | FTS5 BM25 + graph traversal |
+| **Knowledge form** | Implicit (vectors) | Explicit (readable Markdown) |
+| **Auditability** | Low | High (git diff, lint) |
+| **Multi-hop reasoning** | LLM-dependent | Via `related` graph links |
+| **Embedding required** | Yes | No (pure FTS) |
+| **Knowledge accumulation** | None (static index) | Compounds over time |
+| **Token cost at query time** | High (raw chunks passed to LLM) | Low (pre-compiled page) |
+| **Latency** | Higher (ANN + LLM) | Lower (FTS sub-millisecond) |
+| **Infrastructure** | Vector DB required | SQLite only |
+| **Human readability** | No (vectors opaque) | Yes (plain Markdown) |
+| **Best for** | Broad doc Q&A, real-time/live data | Long-term knowledge, agent memory, research |
+
+## When to Use Each
+
+**Choose RAG when:**
+- You need to query a large, dynamic corpus that changes frequently
+- Real-time or near-real-time data ingestion is required
+- Semantic similarity across heterogeneous documents matters most
+- You don't need to maintain or review the knowledge layer
+
+**Choose LLM Wiki when:**
+- Knowledge needs to accumulate and improve over time
+- Auditability and human review of knowledge are important
+- Agents need to search and reason across a curated knowledge base
+- You want to minimize query-time token costs
+- You want offline-capable, zero-infrastructure search
+
+## They Are Not Mutually Exclusive
+
+Many production systems combine both:
+- **RAG** for broad retrieval across large, changing corpora
+- **LLM Wiki** as a curated, high-quality knowledge layer for known domains
+
+Example hybrid: raw ingestion via RAG pipeline → high-confidence results compiled into wiki pages for long-term reuse.
+
+---
+
+See also: [RAG Technologies](/ecosystem/rag-technologies) · [LLM Wiki Technologies](/ecosystem/llm-wiki-technologies)
diff --git a/docs/site/ecosystem/rag-technologies.md b/docs/site/ecosystem/rag-technologies.md
new file mode 100644
index 0000000..848a8d3
--- /dev/null
+++ b/docs/site/ecosystem/rag-technologies.md
@@ -0,0 +1,107 @@
+# RAG Technologies
+
+Core technologies behind the RAG (Retrieval-Augmented Generation) pipeline — from raw document to generated answer.
+
+```
+Raw docs → chunk → embed → store → [query time] → retrieve → LLM → answer
+```
+
+## 1. Document Parsing
+
+| Technology | Representative Tools | Notes |
+|---|---|---|
+| Multi-format parsing | [Unstructured.io](https://unstructured.io), [Docling](https://github.com/DS4SD/docling) (IBM) | Layout-aware extraction from PDF/Word/HTML |
+| PDF parsing | PyMuPDF, PDFPlumber | Low-level text/table/layout extraction |
+| LLM-enhanced parsing | LlamaParse | Cloud API for complex PDFs with tables and figures |
+| PDF → Markdown | [Marker](https://github.com/VikParuchuri/marker), [MinerU](https://github.com/opendatalab/MinerU) | High-fidelity conversion, preserves structure |
+| OCR cloud services | Azure Document Intelligence, AWS Textract | Enterprise-grade form and table recognition |
+
+## 2. Chunking Strategies
+
+| Strategy | Tools | Notes |
+|---|---|---|
+| Fixed-length chunking | LangChain `CharacterTextSplitter` | Simple, breaks semantic boundaries |
+| Recursive splitting | LangChain `RecursiveCharacterTextSplitter` | Respects paragraph → sentence → word hierarchy |
+| Semantic chunking | LlamaIndex `SemanticSplitter` | Detects topic boundaries via embedding similarity |
+| Chapter-tree indexing | [he-wiki-rag](https://github.com/liuhe37186/he-wiki-rag) | Preserves H1/H2/H3 hierarchy + breadcrumb path |
+| Parent-child chunking | LlamaIndex | Retrieve small chunks, return parent as context |
+| Proposition chunking | Custom + LLM | Split into atomic facts — highest precision, expensive |
+| Late Chunking | Jina AI | Embed full document first, then split embeddings |
+| RAPTOR | LlamaIndex | Recursive tree summaries for hierarchical retrieval |
+| Document-tree reasoning | [PageIndex](https://github.com/vectifyai/vectify) | LLM traverses summary tree — no vectors needed, 98.7% on FinanceBench |
+
+## 3. Embedding Models
+
+| Model | Provider | Notes |
+|---|---|---|
+| text-embedding-3-large/small | OpenAI | General-purpose, 3072-dim large variant |
+| Embed v3 | Cohere | Optimized for retrieval, 1024-dim |
+| BGE-M3 | BAAI (open source) | Multilingual, multi-granularity, strong Chinese |
+| E5-Mistral-7B | Microsoft (open source) | High-dim, top MTEB ranking |
+| NV-Embed-v2 | NVIDIA (open source) | MTEB leader among open models |
+| nomic-embed-text | Nomic (open source) | Fully open, local-friendly |
+| jina-embeddings-v3 | Jina AI (open source) | Supports Late Chunking |
+
+## 4. Vector Databases
+
+| Database | Type | Notes |
+|---|---|---|
+| [Qdrant](https://github.com/qdrant/qdrant) | Open source | Rust, high-performance, strong filtering. 100k–10M scale. |
+| [Milvus](https://github.com/milvus-io/milvus) | Open source | Billion-scale distributed. Commercial: Zilliz Cloud. |
+| [Weaviate](https://github.com/weaviate/weaviate) | Open source | Native hybrid search (vector + BM25) |
+| [Chroma](https://github.com/chroma-core/chroma) | Open source | Lightweight, embedded, best for prototyping |
+| [pgvector](https://github.com/pgvector/pgvector) | Extension | Vector search inside PostgreSQL |
+| [LanceDB](https://github.com/lancedb/lancedb) | Open source | Arrow format, embedded, serverless-friendly |
+| Pinecone | Managed SaaS | Serverless, zero ops |
+
+## 5. Retrieval Strategies
+
+| Strategy | Tools | Notes |
+|---|---|---|
+| Dense retrieval | All vector DBs (ANN) | Cosine/dot-product semantic similarity |
+| Sparse retrieval (BM25) | Elasticsearch, OpenSearch, Tantivy | Term-frequency keyword matching |
+| Hybrid retrieval | Weaviate, Qdrant, RRF algorithm | Dense + sparse, merged via Reciprocal Rank Fusion |
+| Graph-augmented retrieval | [GraphRAG](https://github.com/microsoft/graphrag), [LightRAG](https://github.com/HKUDS/LightRAG) | Entity/relation graph for multi-hop reasoning |
+| Vector Graph RAG | Community | Triples vectorized instead of graph DB — 96.3% on HotpotQA |
+| Agentic RAG (A-RAG) | Custom + LLM | Agent autonomously chooses `keyword_search` / `semantic_search` / `chunk_read` tools |
+
+## 6. Query Optimization
+
+| Technique | Tools | Notes |
+|---|---|---|
+| HyDE | LangChain, LlamaIndex | Generate hypothetical answer, use its embedding to retrieve |
+| Multi-query / Query rewriting | LangChain `MultiQueryRetriever` | LLM rewrites to multiple sub-questions |
+| Step-Back Prompting | LangChain | Abstract specific question to general before retrieval |
+| Self-RAG | Research implementation | LLM self-evaluates retrieval quality, decides whether to re-retrieve |
+| CRAG | Research implementation | Falls back to web search for low-confidence retrievals |
+
+## 7. Reranking
+
+| Tool | Notes |
+|---|---|
+| Cohere Rerank | Cross-encoder commercial reranker |
+| BGE Reranker (BAAI) | Open-source cross-encoder, strong multilingual |
+| FlashRank | Lightweight open-source reranker for local deployment |
+
+## 8. Orchestration Frameworks
+
+| Framework | Notes |
+|---|---|
+| [LangChain](https://github.com/langchain-ai/langchain) | Most popular modular RAG pipeline framework |
+| [LlamaIndex](https://github.com/run-llama/llama_index) | Data-centric RAG, richest chunking/retrieval strategies |
+| [Haystack](https://github.com/deepset-ai/haystack) | Enterprise-grade NLP/RAG pipeline |
+| [DSPy](https://github.com/stanfordnlp/dspy) | Programmatic LLM optimization, replaces hand-written prompts |
+
+## 9. Evaluation
+
+| Tool | Notes |
+|---|---|
+| [RAGAS](https://ragas.io) | Most popular RAG eval — faithfulness, context recall, answer relevancy |
+| [DeepEval](https://github.com/confident-ai/deepeval) | Unit-test style RAG evaluation |
+| [TruLens](https://www.trulens.org) | Production monitoring via RAG Triad |
+| LangSmith | LangChain's tracing and eval platform |
+| Key metrics | Faithfulness, Context Recall, Answer Relevancy, MRR, NDCG |
+
+---
+
+See also: [LLM Wiki Technologies](/ecosystem/llm-wiki-technologies) · [Paradigm Comparison](/ecosystem/paradigm-comparison)
diff --git a/docs/site/zh-CN/ecosystem/llm-wiki-technologies.md b/docs/site/zh-CN/ecosystem/llm-wiki-technologies.md
new file mode 100644
index 0000000..3ba4a65
--- /dev/null
+++ b/docs/site/zh-CN/ecosystem/llm-wiki-technologies.md
@@ -0,0 +1,77 @@
+# LLM Wiki 关键技术
+
+LLM Wiki（知识编译范式）背后的核心技术——知识在写入时预编译，查询时直接读取。
+
+```
+原始文档 → LLM 提炼 → Wiki 页面 → [查询时] → FTS / 图遍历 → 读取编译页面 → 答案
+```
+
+> 由 Andrej Karpathy 提出：将 LLM 用作编译器，把原始材料转化为持久化、结构化、可审计的知识层。
+
+## 1. 知识提炼 / 编译
+
+| 技术 | 说明 |
+|---|---|
+| LLM 作为编译器 | LLM 读取原始文档，写入结构化 Wiki 页面——查询时直接读取编译产物。 |
+| 提炼管道 | 原始文档 → LLM 提取关键主张、实体、别名、关联链接 → source-note 页面（WikiLoop 模式）。 |
+| 分层编译 | 优先编译高频文档（Tier 0–3）。Sage Wiki：命中 3 次自动升级，90 天不活跃自动降级。 |
+| 增量更新 | 仅处理变更文档，避免全量重建成本。 |
+| [STORM](https://github.com/stanford-oval/storm)（Stanford OVAL） | 多视角 Wiki 生成 Agent：模拟不同视角提问 → 检索 → 生成层级大纲 → 撰写带引用的完整词条。 |
+| Co-STORM | STORM 的协作变体，研究过程中构建动态知识图谱指导编译方向。 |
+| Agentic 编译循环 | Agent 循环：检索 → 草稿 → 评估 → 补充检索 → 再生成，直到知识稳定。 |
+
+## 2. 知识表示
+
+| 技术 | 说明 |
+|---|---|
+| 结构化 Markdown | 所有知识以纯文本 Markdown 存储——人类可读，可 git diff，可审计。 |
+| Source-note 页面 | 每个原始文档对应一个提炼笔记，包含 `key_claims`、实体标注 `【实体\|类型】`、`related_to` / `supports` / `contradicts` 链接。 |
+| Concept 页面 | 对某一概念或方法论的跨文档综合描述。 |
+| Comparison 页面 | 方案、工具或范式的横向对比。 |
+| Decision 页面 | 技术决策记录（ADR 格式）：背景 + 选项 + 决策 + 影响。 |
+| 本体图谱（Ontology Graph） | 编译过程中构建的类型化实体-关系图谱。Sage Wiki 内置 8 种关系类型（`implements`、`contradicts`、`trades_off` 等）。 |
+| Schema / 模板 | 指导 LLM 编译风格的写作规则和页面模板，每个 KB 可自定义。 |
+
+## 3. 索引与搜索
+
+| 技术 | 说明 |
+|---|---|
+| SQLite FTS5 + BM25 | 核心搜索引擎——无需向量模型，亚毫秒级全文检索。WikiLoop、Sage Wiki、TreeSearch 均采用。 |
+| 别名扩展 | 关键术语索引时内嵌别名和跨语言等价词，最大化 FTS 召回率。 |
+| 图遍历 | `related_to` 链接实现多跳导航（类似 Wiki 页面链接）。搜索结果进行 BFS 扩展。 |
+| 混合检索（FTS + 向量 + 图） | Sage Wiki：FTS5（411µs）+ 向量（81ms）+ 本体图谱（1µs），通过 RRF 融合。 |
+| 章节树索引 | 保留文档 H1/H2/H3 层级——平铺切块的结构保留替代方案。 |
+
+## 4. 知识质量与维护
+
+| 技术 | 说明 |
+|---|---|
+| Git 版本控制 | 所有 Wiki 页面在 Git 中——完整历史、diff、blame，知识变更完全可审计。 |
+| Lint / 健康检查 | 验证 frontmatter、断开的 source 链接、缺失引用。`wikiloop lint`。 |
+| 冲突检测 | `contradicts` 链接将来源间的分歧暴露出来，供人工审核。 |
+| Draft 暂存 | 来源少于 2 个的页面隔离到 `_draft/`——验证补充后才进入索引。 |
+| 知识空白分析 | `wikiloop synthesize --gaps` 识别覆盖不足的主题。 |
+| 实体去重 / 解析 | 识别同一概念的不同表述并合并为单一节点。 |
+
+## 5. Agent 接口（MCP）
+
+| 技术 | 说明 |
+|---|---|
+| [MCP 协议](https://modelcontextprotocol.io) | Model Context Protocol——Anthropic 开放标准，向 AI Agent 暴露工具/资源。支持 stdio + HTTP 双传输。 |
+| `kb_search` | FTS 关键词搜索，返回带 `related` 链接的排序结果，用于图谱导航。 |
+| `kb_page` | 通过 ID 获取完整页面内容，支持批量（最多 5 个 ID）或 `full=true` 获取不截断文本。 |
+| MCP Resources | 只读资源（URI 形式）：Wiki 页面、图谱 Schema、原始文档。 |
+| 迭代搜索模式 | Agent 从不同角度发出多次查询，跟随 `related` 链接，自行综合答案。 |
+| Sage Wiki MCP 工具 | 17 个工具：6 读、9 写、2 复合——Agent 可直接写入和编译知识。 |
+
+## 6. 文件转换（输入层）
+
+| 技术 | 说明 |
+|---|---|
+| markitdown（Microsoft） | 将 PDF、Word、Excel、PPT、HTML 转为 Markdown 后再提炼。 |
+| `raw/converted/` 模式 | Agent 提取的内容直接写入此目录——跳过转换，直接进入提炼流程。 |
+| 文件监听器（Watcher） | 自动检测 `raw/` 下的新文件/变更文件，触发转换 → 提炼 → 索引全流程。 |
+
+---
+
+另见：[RAG 关键技术](/zh-CN/ecosystem/rag-technologies) · [范式对比](/zh-CN/ecosystem/paradigm-comparison)
diff --git a/docs/site/zh-CN/ecosystem/paradigm-comparison.md b/docs/site/zh-CN/ecosystem/paradigm-comparison.md
new file mode 100644
index 0000000..bc089ea
--- /dev/null
+++ b/docs/site/zh-CN/ecosystem/paradigm-comparison.md
@@ -0,0 +1,65 @@
+# 范式对比：RAG vs LLM Wiki
+
+构建 AI 知识基础设施的两种根本不同的路径。
+
+## 管道对比
+
+```
+RAG 管道
+────────────────────────────────────────────────────────
+原始文档 → 切块 → 嵌入 → 向量存储
+                                    ↘ [查询时]
+                         查询 → ANN 搜索 → LLM → 答案
+
+
+LLM Wiki（知识编译范式）
+────────────────────────────────────────────────────────
+原始文档 → LLM 编译 → 结构化 Wiki 页面 → FTS 索引
+                                                    ↘ [查询时]
+                              查询 → FTS / 图遍历 → 读取页面 → 答案
+```
+
+## 多维对比
+
+| 维度 | RAG | LLM Wiki |
+|---|---|---|
+| **核心操作** | 查询时检索 | 写入时编译 |
+| **存储** | 向量数据库 + 原始 chunk | 结构化 Markdown + SQLite FTS |
+| **搜索方式** | ANN 相似度搜索 | FTS5 BM25 + 图遍历 |
+| **知识形态** | 隐式（向量） | 显式（可读 Markdown） |
+| **可审计性** | 低 | 高（git diff、lint） |
+| **多跳推理** | 依赖 LLM | 通过 `related` 图链接 |
+| **是否需要 Embedding** | 是 | 否（纯 FTS） |
+| **知识积累** | 无（静态索引） | 随时间复利增长 |
+| **查询时 Token 成本** | 高（原始 chunk 传给 LLM） | 低（预编译页面） |
+| **延迟** | 较高（ANN + LLM） | 较低（FTS 亚毫秒） |
+| **基础设施** | 需要向量数据库 | 仅需 SQLite |
+| **人类可读性** | 否（向量不透明） | 是（纯 Markdown） |
+| **最适合场景** | 宽泛文档问答、实时/动态数据 | 长期知识管理、Agent 记忆、研究 |
+
+## 如何选择
+
+**选 RAG 的场景：**
+- 需要查询频繁变化的大规模语料
+- 需要实时或近实时数据摄入
+- 异构文档间的语义相似度匹配最重要
+- 不需要维护或审查知识层
+
+**选 LLM Wiki 的场景：**
+- 知识需要随时间积累和改进
+- 知识的可审计性和人工审核很重要
+- Agent 需要在精心策展的知识库上搜索和推理
+- 希望最小化查询时的 Token 成本
+- 希望离线可用、零基础设施依赖
+
+## 两者并不互斥
+
+许多生产系统将两者结合使用：
+- **RAG** 负责对大规模、动态变化语料的宽泛检索
+- **LLM Wiki** 作为精心策展的高质量知识层，覆盖已知领域
+
+典型混合模式：RAG 管道宽泛摄入 → 高置信度结果编译为 Wiki 页面供长期复用。
+
+---
+
+另见：[RAG 关键技术](/zh-CN/ecosystem/rag-technologies) · [LLM Wiki 关键技术](/zh-CN/ecosystem/llm-wiki-technologies)
diff --git a/docs/site/zh-CN/ecosystem/rag-technologies.md b/docs/site/zh-CN/ecosystem/rag-technologies.md
new file mode 100644
index 0000000..01121a0
--- /dev/null
+++ b/docs/site/zh-CN/ecosystem/rag-technologies.md
@@ -0,0 +1,107 @@
+# RAG 关键技术
+
+RAG（检索增强生成）管道背后的核心技术——从原始文档到生成答案的全链路。
+
+```
+原始文档 → 切块 → 嵌入 → 存储 → [查询时] → 检索 → LLM → 答案
+```
+
+## 1. 文档解析层
+
+| 技术 | 代表工具 | 说明 |
+|---|---|---|
+| 多格式文档解析 | [Unstructured.io](https://unstructured.io)、[Docling](https://github.com/DS4SD/docling)（IBM） | 版面感知的 PDF/Word/HTML 多格式提取 |
+| PDF 底层解析 | PyMuPDF、PDFPlumber | 文本/表格/布局结构提取 |
+| LLM 增强解析 | LlamaParse | 处理含复杂表格和图形 PDF 的云端 API |
+| PDF 转 Markdown | [Marker](https://github.com/VikParuchuri/marker)、[MinerU](https://github.com/opendatalab/MinerU) | 高保真格式转换，保留文档结构 |
+| OCR 云服务 | Azure Document Intelligence、AWS Textract | 企业级表单和表格识别 |
+
+## 2. 文本切块策略
+
+| 策略 | 工具 | 说明 |
+|---|---|---|
+| 固定长度切块 | LangChain `CharacterTextSplitter` | 简单，但会破坏语义边界 |
+| 递归字符切割 | LangChain `RecursiveCharacterTextSplitter` | 按段落→句子→单词层级递归切割 |
+| 语义切块 | LlamaIndex `SemanticSplitter` | 通过 Embedding 相似度检测主题边界 |
+| 章节树索引 | [he-wiki-rag](https://github.com/liuhe37186/he-wiki-rag) | 保留 H1/H2/H3 层级和面包屑路径 |
+| 父子分块 | LlamaIndex | 检索小粒度子块，返回父块作为上下文 |
+| 命题切块 | 自定义 + LLM | 拆分为原子事实——精度最高，代价最大 |
+| Late Chunking | Jina AI | 先嵌入完整文档再切分 Embedding |
+| RAPTOR | LlamaIndex | 递归树状摘要，构建层次化检索结构 |
+| 文档树推理 | [PageIndex](https://github.com/vectifyai/vectify) | LLM 遍历摘要树检索，无需向量，FinanceBench 98.7% |
+
+## 3. Embedding 模型
+
+| 模型 | 提供方 | 说明 |
+|---|---|---|
+| text-embedding-3-large/small | OpenAI | 通用嵌入，large 版 3072 维 |
+| Embed v3 | Cohere | 专为检索优化，1024 维 |
+| BGE-M3 | BAAI（开源） | 多语言多粒度，中文表现优异 |
+| E5-Mistral-7B | Microsoft（开源） | 高维度，MTEB 排名靠前 |
+| NV-Embed-v2 | NVIDIA（开源） | 开源模型中 MTEB 领先 |
+| nomic-embed-text | Nomic（开源） | 完全开源，本地友好 |
+| jina-embeddings-v3 | Jina AI（开源） | 支持 Late Chunking |
+
+## 4. 向量数据库
+
+| 数据库 | 类型 | 说明 |
+|---|---|---|
+| [Qdrant](https://github.com/qdrant/qdrant) | 开源 | Rust 实现，高性能，强过滤，适合 10 万~千万规模 |
+| [Milvus](https://github.com/milvus-io/milvus) | 开源 | 十亿级分布式，商业版：Zilliz Cloud |
+| [Weaviate](https://github.com/weaviate/weaviate) | 开源 | 原生混合检索（向量 + BM25） |
+| [Chroma](https://github.com/chroma-core/chroma) | 开源 | 轻量嵌入式，原型开发首选 |
+| [pgvector](https://github.com/pgvector/pgvector) | 扩展 | PostgreSQL 原生向量搜索扩展 |
+| [LanceDB](https://github.com/lancedb/lancedb) | 开源 | Arrow 格式，嵌入式，无服务器友好 |
+| Pinecone | 托管 SaaS | Serverless，零运维 |
+
+## 5. 检索策略
+
+| 策略 | 工具 | 说明 |
+|---|---|---|
+| 密集检索 | 各向量数据库 ANN | 余弦/点积语义相似度近似最近邻搜索 |
+| 稀疏检索（BM25） | Elasticsearch、OpenSearch、Tantivy | 基于词频统计的关键词匹配 |
+| 混合检索 | Weaviate、Qdrant、RRF 算法 | 密集 + 稀疏，通过 RRF 倒数排名融合 |
+| 图增强检索 | [GraphRAG](https://github.com/microsoft/graphrag)、[LightRAG](https://github.com/HKUDS/LightRAG) | 实体/关系图谱支持多跳推理 |
+| Vector Graph RAG | 社区 | 三元组向量化替代图数据库，HotpotQA 96.3% |
+| Agentic RAG（A-RAG） | 自定义 + LLM | Agent 自主选择 `keyword_search` / `semantic_search` / `chunk_read` 工具 |
+
+## 6. 查询优化
+
+| 技术 | 工具 | 说明 |
+|---|---|---|
+| HyDE | LangChain、LlamaIndex | 生成假设性答案，用其 Embedding 做检索 |
+| Multi-Query 查询改写 | LangChain `MultiQueryRetriever` | LLM 将原问题改写为多个子问题扩大召回 |
+| Step-Back Prompting | LangChain | 将具体问题抽象为通用问题再检索 |
+| Self-RAG | 论文实现 | LLM 自评估检索质量，决定是否继续检索 |
+| CRAG | 论文实现 | 低置信度时回退到网络搜索 |
+
+## 7. Rerank 精排
+
+| 工具 | 说明 |
+|---|---|
+| Cohere Rerank | 交叉编码器商业精排模型 |
+| BGE Reranker（BAAI） | 开源交叉编码器，多语言表现强 |
+| FlashRank | 轻量开源精排库，适合本地部署 |
+
+## 8. 编排框架
+
+| 框架 | 说明 |
+|---|---|
+| [LangChain](https://github.com/langchain-ai/langchain) | 最流行的模块化 RAG 管道框架 |
+| [LlamaIndex](https://github.com/run-llama/llama_index) | 数据中心型 RAG，最丰富的切块/检索策略 |
+| [Haystack](https://github.com/deepset-ai/haystack) | 企业级 NLP/RAG 管道框架 |
+| [DSPy](https://github.com/stanfordnlp/dspy) | 通过编程方式优化 LLM 管道，替代手写 Prompt |
+
+## 9. 评估框架
+
+| 工具 | 说明 |
+|---|---|
+| [RAGAS](https://ragas.io) | 最流行的 RAG 评估框架——忠实度、上下文召回、答案相关性 |
+| [DeepEval](https://github.com/confident-ai/deepeval) | 单元测试风格的 RAG 评估库 |
+| [TruLens](https://www.trulens.org) | 基于 RAG Triad 的生产监控工具 |
+| LangSmith | LangChain 官方链路追踪与评估平台 |
+| 核心指标 | Faithfulness、Context Recall、Answer Relevancy、MRR、NDCG |
+
+---
+
+另见：[LLM Wiki 关键技术](/zh-CN/ecosystem/llm-wiki-technologies) · [范式对比](/zh-CN/ecosystem/paradigm-comparison)