rag-mcp-server

Your documents and knowledge graph as tools that any AI assistant can call, over the Model Context Protocol.

The client (Claude Desktop, Claude Code, an IDE, or your own agent) decides when to search passages, when to walk the graph, and when to inspect an entity. This server only retrieves. It returns cited evidence (document, page, section, or ordered fact chains) and never answers on the model's behalf.

Run it

pip install "rag-mcp-server @ git+https://github.com/gopalsharma1615-ai/rag-mcp-server"

layout-rag ingest docs/*.pdf --index ./index          # documents (see layout-rag)
graphwalk-rag build docs/ --workdir ./graph           # graph (see graphwalk-rag; price it with --dry-run)

export LAYOUT_RAG_INDEX=./index GRAPH_WORKDIR=./graph
export GRAPH_EMBED_MODEL=text-embedding-3-small        # the model that embedded the graph
export LLM_MODEL=gpt-4o-mini                           # keyword extraction for the walks (or LLM_BASE_URL for vLLM)

claude mcp add rag-tools -- rag-mcp-server             # Claude Code (stdio)
rag-mcp-server --http 8765                             # or streamable HTTP for remote clients

Claude Desktop (claude_desktop_config.json):

{"mcpServers": {"rag-tools": {"command": "rag-mcp-server",
  "env": {"LAYOUT_RAG_INDEX": "/path/index", "GRAPH_WORKDIR": "/path/graph"}}}}

Either backend is optional. Without GRAPH_WORKDIR only search_documents works, and vice versa. Indexes and models load lazily on first use.

Related MCP server: Provenance-Aware Retrieval MCP Server

Measured latency

Warm latency on a CPU-only VM (8 threads), per tool call, from python bench/latency.py:

These measurements drove a design decision: reranking is opt-in (rerank=False by default, pool set by RAG_RERANK_POOL, default 20). A 568M-parameter cross-encoder over 512-token chunks is impractical on CPU for an interactive tool: even 20 candidates take about 40 s. Use a GPU, a smaller cross-encoder (RAG_RERANKER), or leave it off. The graph tools' latency is dominated by the keyword-extraction LLM call and query embedding, not by the walk itself: the traversal takes milliseconds (see graphwalk-rag).

Evaluation plan

The agent-vs-pipeline comparison is planned and not yet run: an LLM client choosing these tools vs a fixed single-pass pipeline on HotpotQA multi-hop, measuring answer accuracy, tool-selection accuracy, calls per question, latency and $/question. It will live in research-agent.

Development

pip install -e ".[dev]" && pytest -q   # tool tests on a tiny index + graph, and a real stdio round-trip

MIT licence. Built by Gopal Sharma.