Enterprise Internal Knowledge Base — Production-Ready RAG + MCP

A public Retrieval-Augmented Generation pipeline exposed as an MCP server. Sample content from Veterans Affairs education manuals.

The repo implements evaluation, observability, and structure-aware ingestion. Cost/latency tuning, tenant-level access control, and other production concerns are discussed in the article linked below.

📖 Full writeup on Medium: Enterprise Internal Knowledge Base RAG MCP: POC-to-Production

Why this exists

RAG demos tend to focus on the quality of the retrieval pipeline, without recognizing that production RAG fails on the next ten steps: prompt or model changes that pass code review but tank answer quality, cost and latency drift that cannot be traced to specific queries, cross-tenant leakage that only surfaces in audit. This repo shows what catching them looks like in practice.

The corpus is public (VA Education manuals — 238 documents, 9,000+ chunks) so anyone can clone, run, and adapt the pipeline.

Quickstart

git clone https://github.com/kimsb2429/internal-knowledge-base
cd internal-knowledge-base

# 1. Start Postgres + pgvector
docker compose up -d

# 2. Python env + dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 3. Restore corpus fixture (~2 min — 238 docs + 9k chunks pre-embedded)
docker exec -i ikb_pgvector pg_restore -U ikb -d ikb < evals/fixture_v1.dump

# 4. Smoke-test the MCP server
python scripts/test_mcp_server.py     # 7/7 tests pass

# 5. Start the MCP server (stdio transport)
python scripts/mcp_server.py

Consuming from Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "ikb": {
      "command": "python",
      "args": ["/absolute/path/to/internal-knowledge-base/scripts/mcp_server.py"]
    }
  }
}

Then ask Claude things like "What RPO handles GI Bill claims in Texas?" — the MCP server returns ranked chunks with citations.

Architecture

Ingestion (one-time per corpus):

graph LR
    A[KnowVA crawler<br/>HTML + PDF] --> B[Source-specific<br/>preprocessor]
    B --> C[Structure-aware<br/>chunker]
    C --> D[mxbai-embed-large<br/>local, 1024-dim]
    D --> E[(pgvector)]
    F[Anthropic Contextual<br/>Retrieval] -.-> E
    E -.-> F
    style E fill:#e1f5fe

Query (per MCP tool call):

graph LR
    A[Claude Desktop<br/>MCP client] --> B[FastMCP server]
    B --> C[pgvector top-K]
    C --> D[Reranker<br/>mxbai or FlashRank]
    D --> E[Claude Sonnet<br/>generation]
    E --> A
    E --> F[Langfuse trace]
    style F fill:#fff9c4

Stack:

• Vector store: Postgres + pgvector (Docker, port 5433); content_tsv GIN index for hybrid-ready
• Embeddings: mxbai-embed-large (1024 dims, local via sentence-transformers) — $0 API cost
• Reranker: mxbai-rerank-base-v2 (full eval) / FlashRank MiniLM (CI fast mode, 22M ONNX, ~2s/query)
• Generation: Claude Sonnet
• MCP server: FastMCP 3.2.4 — Tools (query), Resources (document://{source_id}), Prompts (cite_from_chunks)
• Observability: Langfuse Cloud, per-trace public sharing
• Eval: DeepEval + 110-query golden set + GitHub Actions merge gate

Eval scores

Full 110-question golden set, contextualized chunks + reranker:

🔗 Live Langfuse trace (public, no login).

Notable result: Anthropic's Contextual Retrieval pattern produced modest lift on top of reranking (+4.8pp AnsRel, +4.1pp CtxPrec) at this scale — well short of the +35% recall their published numbers suggested. Rep

... View full README on GitHub