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 in Towards AI: 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

bash

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:

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):

mermaid

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):

mermaid

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:

Metric	Score
Faithfulness	0.95
Answer Relevance	0.91
Context Precision	0.61
Context Recall	0.52
Context Relevance	0.56

🔗 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. Reported as found; juiced numbers would defeat the point.

Eval-in-CI as a merge gate

Every PR runs the golden set in fast mode (FlashRank reranker, ~3-4 min wall, $0.30 in Sonnet calls) against a fixture DB. PRs that regress more than ±5pp on top1/topk/keyword_recall, or +10pp on idk_rate, are blocked.

Forever-artifact: PR #5 — a deliberate failing-then-passing PR. Red CI catches a 20pp top1 regression; green CI confirms the fix. The Actions tab is the proof.

Workflow: .github/workflows/eval-gate.yml.

What this repo doesn't cover

A few production-shape items are seams, not implementations:

Multi-tenant scoping — auth_context parameter present on every MCP tool, typed, currently unused (labels the SSO/ACL seam)
Ingestion concurrency — single-threaded chunker + embedder; production would use a modulus-distributed worker pool
Hybrid search wiring — content_tsv GIN index is live; BM25 + RRF fusion at query time stays a post-launch addition

The writeup linked above covers these topics.

Repo layout

Code

docs/                    Research, evidence base, deep-dives
data/                    Crawled corpus + golden query set
scripts/
  crawl_knowva.py            eGain v11 API crawler
  enrich_metadata.py         Headings, ACL, authority tier, content_category
  knowva_preprocess.py       Source-specific HTML normalization
  chunk_documents.py         Structure-aware splitter (preserves table colspan/rowspan)
  embed_and_store.py         mxbai-embed-large → pgvector
  contextualize_chunks.py    Anthropic Batches API for Contextual Retrieval
  rerank.py                  mxbai-rerank + FlashRank
  retrieve.py / generate.py  RAG path
  mcp_server.py              FastMCP exposure
  run_eval.py / score_eval.py / check_regression.py   Eval harness + CI gate
evals/                   Fixture DB dump + baseline JSON
.github/workflows/       eval-gate.yml — merge-gate workflow

Reproducing from raw corpus (~30 min)

Each script is idempotent and resume-safe.