Agentic RAG MCP

A multi-agent Retrieval-Augmented Generation system exposed as an MCP server. Ask a question and a LangGraph pipeline plans the retrieval, pulls evidence from a pgvector knowledge base, optionally augments it with live web research, drafts a cited answer, and then self-critiques it for grounding — revising until the answer is supported by the sources.

It plugs into any MCP client (Claude Code/Desktop, Cursor, Windsurf, …) as three tools: ingest, ask, and search.

Why this design? A bare RAG endpoint is easy to copy; a multi-agent system that verifies its own answers and ships as an MCP server is not. The architecture is the moat — "easy to buy, hard to replicate."

Architecture

flowchart LR
    Q([Question]) --> P[🧭 Planner<br/>plan + search queries]
    P --> R[📚 Retriever<br/>pgvector top-k]
    R --> W[🌐 Web Researcher<br/>Firecrawl • optional]
    W --> S[✍️ Synthesizer<br/>cited answer]
    S --> C{🔎 Critic<br/>grounded?}
    C -- needs revision --> S
    C -- grounded --> A([Answer + citations])

    subgraph Stores
      DB[(Supabase<br/>pgvector)]
    end
    R <-->|cosine search| DB

    classDef agent fill:#1e293b,stroke:#7C3AED,color:#e2e8f0;
    class P,R,W,S,C agent;

Related MCP server: Corpus-KB

MCP tools

Quickstart

# 1. Install (Python 3.10+)
uv venv && uv pip install -e ".[dev]"     # or: pip install -e ".[dev]"

# 2. Configure
cp .env.example .env                       # fill in ANTHROPIC_API_KEY, VOYAGE_API_KEY, DATABASE_URL

# 3. Create the vector table (Supabase SQL editor or psql)
psql "$DATABASE_URL" -f sql/schema.sql

# 4. Run the MCP server (stdio by default)
agentic-rag-mcp

Connect it to Claude Code

claude mcp add agentic-rag -s user \
  --env ANTHROPIC_API_KEY=sk-ant-... \
  --env VOYAGE_API_KEY=pa-... \
  --env DATABASE_URL=postgresql://... \
  -- agentic-rag-mcp

Then, from the client: "ingest https://example.com/docs" → "ask: how do I configure X?".

How it works

1. Plan — Claude turns the question into a short plan + 1–5 search queries.

2. Retrieve — each query is embedded (Voyage voyage-3.5) and matched against pgvector by cosine distance; results are de-duplicated and ranked.

3. Research — if FIRECRAWL_API_KEY is set, live web results are added to the context.

4. Synthesize — Claude writes an answer grounded only in the numbered context, citing each claim as [n].

5. Critique — a strict fact-checker pass decides whether the answer is fully supported. If not (and revisions remain), it loops back to the synthesizer with feedback.

Configurable via env: RAG_MODEL, RAG_TOP_K, RAG_MAX_REVISIONS, RAG_EMBED_MODEL.

Evaluation — the CI gate

📊 Live eval dashboard: enached134-ctrl.github.io/agentic-rag-mcp — the golden dataset and the latest green run (20/20 passing), in one screen.

Answer quality is measured with promptfoo on a golden dataset (evals/golden.yaml — 20 seed cases: answerable / refusal / adversarial, written against a committed corpus) and enforced in CI on every push: the evals job spins up a pgvector service container, seeds it with evals/corpus/, and runs every case through the real pipeline — planner, retriever, synthesizer, self-critique. Nothing is mocked. A regression fails the build before it can reach a user.

Scored dimensions: citation presence (deterministic) · groundedness (LLM-as-judge) · refusal correctness (LLM-as-judge) · latency (threshold).

python evals/seed.py --schema --reset   # seed the corpus into your vector store
make eval                               # run the suite locally

See evals/ for the corpus, the golden dataset, and the regression-capture rule: every real-world failure becomes a new golden case, so no bug gets fixed twice.

The CI gate activates when the ANTHROPIC_API_KEY and VOYAGE_API_KEY repository secrets are configured; without them (e.g. on forks) the job skips with a visible notice.

Observability

Opt-in OpenTelemetry tracing to Arize Phoenix:

pip install -e ".[trace]"
phoenix serve                          # local Phoenix UI on :6006
PHOENIX_ENABLED=1 agentic-rag-mcp

Every ask run appears as a full trace — LangGraph node spans (plan → retrieve → research → synthesize → critique) plus every Claude call with token usage and latency per span. Point PHOENIX_COLLECTOR_ENDPOINT at a hosted collector to ship traces off-box.

Deploy

Containerised and ready for Railway (HTTP transport):

railway up        # uses Dockerfile + railway.json; set RAG_TRANSPORT=http

Expose RAG_HTTP_PORT and connect over --transport http. A cloudflared tunnel works for local demos.

Kubernetes

Production-shaped manifests — readiness/liveness probes, resource limits, secret-driven env — live in deploy/k8s/, including a kind-based local smoke test walkthrough.

Project layout

src/agentic_rag_mcp/
  config.py      # env-driven settings
  llm.py         # Anthropic (Claude) helper — adaptive thinking, JSON parsing
  embeddings.py  # Voyage embeddings
  store.py       # pgvector store (psycopg)
  web.py         # Firecrawl web research (optional)
  ingest.py      # chunking + ingestion
  state.py       # LangGraph state
  nodes.py       # planner / retriever / researcher / synthesizer / critic
  graph.py       # graph assembly
  tracing.py     # opt-in OpenTelemetry → Arize Phoenix
  server.py      # FastMCP server (ingest / ask / search)
sql/schema.sql   # pgvector schema
evals/           # golden dataset + corpus + promptfoo suite (runs in CI)
deploy/k8s/      # Kubernetes manifests + kind smoke test

License

MIT — see LICENSE.