graph-rag

A local-first Graph RAG knowledge base for coding agents, served over MCP.

graph-rag ingests the heterogeneous stuff a coding agent needs to reason about — service docs (PDF), internal Markdown, your source code, and YAML policy files (Checkov) — into a single Neo4j knowledge graph, and exposes it to the agent over an MCP server: hybrid (vector + full-text) search, table-of-contents navigation, exact policy lookup, code-centrality ranking, graph traversal, and the agent's own persistent working memory.

It runs entirely on your machine. The default embedding model is local, so ingestion needs no API key and works offline.

Why

Plain vector RAG loses structure: it can't tell you which section a chunk came from, what calls a function, or which policy applies to aws_db_instance. graph-rag keeps those relationships as graph edges, so an agent can both search semantically and traverse — "find the retry section, then show me its parent chapter", "rank this codebase's most-depended-upon functions", "give me the Checkov rules for this resource type". It also gives the agent a place to remember decisions and recall them in a later session.

Demo

Install

The PyPI package is grag-mcp (the name graph-rag was taken); it installs a grag-mcp command. No clone needed — run it straight with uv:

uvx grag-mcp --help                       # one-off, no install
uvx 'grag-mcp[pdf]' serve-mcp --stdio     # with PDF ingestion support

or install the grag-mcp command onto your PATH:

uv tool install 'grag-mcp[pdf]'    # or: pipx install 'grag-mcp[pdf]'

You still need a Neo4j instance (APOC + GDS plugins) reachable at NEO4J_URI / NEO4J_USER / NEO4J_PASSWORD — see docker-compose.yml for a ready-made one. The [pdf] extra pulls in PyMuPDF (AGPL-licensed); the [java] extra pulls in tree-sitter for .java files, [js] for .js / .mjs / .cjs / .jsx / .ts / .tsx, [sql] pulls in sqlglot for .sql schema DDL and procedural code (PL/SQL, PL/pgSQL, T-SQL — also .pks / .pkb / .prc / .fnc / .trg), and [css] pulls in tree-sitter for .css / .scss / .sass / .less. Leave off any you don't ingest (uv tool install 'grag-mcp[pdf,java,js,sql,css]' for all).

On Linux, pass --torch-backend=cpu (uvx --torch-backend=cpu …) unless you want the multi-gigabyte CUDA build of PyTorch — the embedding model runs on CPU.

A prebuilt runtime image (linux/amd64 + linux/arm64, embedding model baked in) is published on each release:

docker pull ghcr.io/tmustafiz/graph-rag:latest

Quickstart (from a clone)

Requires uv and Docker.

cp .env.example .env        # adjust NEO4J_PASSWORD if you like
make install                # uv sync --all-extras
make fetch-model            # download the local embedding model (~87 MB)
make up                     # start Neo4j (Docker)
make apply-schema           # constraints + full-text + vector indexes
make ingest INGEST_PATH=examples/checkov-policies   # or point at your own docs
make mcp-serve              # MCP server on http://127.0.0.1:8765/mcp

graph-rag ships no document corpus — you bring the files to ingest. examples/ holds a few small samples to try the tooling against; everything else is yours.

Neo4j Browser: http://localhost:7474 (neo4j / your NEO4J_PASSWORD).

Or run everything (Neo4j + MCP server) with Compose:

docker compose up -d

Other targets: make down, make lint, make format, make test, make eval.

Connect an agent

The MCP server speaks Streamable HTTP at http://127.0.0.1:8765/mcp. .mcp.json at the repo root already registers it for this project.

Claude Code

claude mcp add graph-rag --transport http http://127.0.0.1:8765/mcp

Claude Desktop / Cursor / Windsurf / VS Code — add to the MCP config:

{
  "mcpServers": {
    "graph-rag": { "type": "http", "url": "http://127.0.0.1:8765/mcp" }
  }
}

Set MCP_AUTH_TOKEN in .env to require a bearer token (defense in depth; the server is bound to 127.0.0.1 regardless — see SECURITY.md).

stdio transport

For clients that launch the server as a subprocess instead of connecting over HTTP, run grag-mcp serve-mcp --stdio — no port, no auth token, no POST /ingest. Point the client's command at it:

{
  "mcpServers": {
    "graph-rag": { "command": "grag-mcp", "args": ["serve-mcp", "--stdio"] }
  }
}

Use uv run grag-mcp … (or an absolute path to the entry point) as the command if grag-mcp isn't on the client's PATH. Neo4j still has to be reachable at NEO4J_URI.

MCP tools

Ingesting your own content

grag-mcp ingest <path> takes a file or a directory (recursed), parses whichever of PDF / Markdown / Python / Java / JS / TS / SQL / CSS / YAML it finds, and upserts into the graph. Re-running is cheap: a file whose content hash is unchanged since the last ingest is skipped entirely, and re-ingesting a changed file removes any Section / Chunk / CodeEntity / PolicyRule it no longer produces.

uv run grag-mcp ingest src/graph_rag           # this repo's own source
uv run grag-mcp ingest path/to/docs            # a whole directory
uv run grag-mcp ingest some/file.py --dry-run  # preview, no writes
uv run grag-mcp ingest src/graph_rag --watch   # re-ingest on every change

A file that fails to parse/embed/write is reported and skipped rather than aborting the batch — see docs/operations.md.

Ingestion is also reachable over plain HTTP while serve-mcp / docker compose up is running, for triggering from CI or a pre-commit hook without an MCP client:

curl -X POST http://127.0.0.1:8765/ingest \
  -H "Content-Type: application/json" \
  -d '{"path": "src/graph_rag", "dry_run": false}'

Code centrality (PageRank)

grag-mcp compute-centrality runs GDS PageRank over the CodeEntity CALLS/IMPORTS graph, writing each entity's score to CodeEntity.pagerank — a heavily called/imported entity ranks higher, surfacing what's most central (and riskiest to change) in an ingested codebase. Exposed via get_central_code_entities. Needs source code already ingested and the graph-data-science Neo4j plugin (enabled in docker-compose.yml):

uv run grag-mcp ingest src/graph_rag
uv run grag-mcp compute-centrality   # re-run after ingesting code changes

Offline embedding model

The MCP server and ingestion embed with sentence-transformers/all-MiniLM-L6-v2 (Apache-2.0). make fetch-model downloads just the PyTorch + tokenizer files (~87 MB) into models/all-MiniLM-L6-v2/ — see scripts/fetch_model.py. The Docker image bakes the same files in at /opt/models/all-MiniLM-L6-v2, so docker compose up needs no network for embeddings.

SentenceTransformerEmbedder resolves the model in this order: the GRAG_EMBEDDING_MODEL env var (a local directory or a Hub repo id), the copy baked into the image, the models/all-MiniLM-L6-v2/ folder in a checkout, and finally the Hub repo id — the only branch that needs huggingface.co.

Hosted embedding backends (optional)

The local model is the default and needs no API key. To embed with a hosted provider instead, set GRAG_EMBEDDING_PROVIDER to one of openai, ollama, voyage, cohere, or gemini (anything else, or unset, keeps the local model). Each backend is a thin REST call over httpx — no provider SDKs are pulled in.

• GRAG_EMBEDDING_MODEL overrides the model id.
• GRAG_EMBEDDING_API_BASE overrides the endpoint — point openai at any OpenAI-compatible gateway (vLLM, LM Studio, a proxy), or ollama at a non-local host.
• The Neo4j vector index is created at EMBEDDING_DIMENSIONS (384 for the local model). A hosted model with a different width means setting EMBEDDING_DIMENSIONS, re-running grag-mcp apply-schema, and re-ingesting. build_embedder() probes the provider once at startup and refuses to run on a mismatch rather than corrupting the index mid-ingest.
• The Embedder interface has no query-vs-document distinction, so Cohere and Voyage calls always use the document input type.

Reranking (optional)

Hybrid search shortlists candidates with a bi-encoder (fast, compares independently computed vectors). Setting GRAG_RERANK=1 adds a second stage: a cross-encoder (cross-encoder/ms-marco-MiniLM-L-6-v2, Apache-2.0) that reads the query and each shortlisted document together and re-scores them directly — more accurate, but only affordable over the ~20 candidates hybrid search already narrowed to. It applies to search, search_code, and search_policies.

Off by default. The model is not baked into the Docker image, and nothing downloads it implicitly. Make it resolvable in one of two ways:

• make fetch-reranker — vendors it into models/ms-marco-MiniLM-L-6-v2/ (also picked up at /opt/models/ms-marco-MiniLM-L-6-v2 in a container);
• GRAG_RERANK_MODEL=<path> — a local directory, or a Hub repo id such as cross-encoder/ms-marco-MiniLM-L-6-v2, which is the explicit opt-in to an online pull.

With GRAG_RERANK=1 set and neither in place, the server fails at startup with a message naming these fixes — not at the first query.

Reranking reorders the shortlist by the cross-encoder score, with the fused score breaking ties. Each hit keeps its score (the [0, 1] fused value, unchanged) and gains a rerank_score (the raw cross-encoder logit — unbounded, can be negative — null when reranking is off). The reranker only sees the vector shortlist, so a hit that full-text alone would surface is not rescued by it.

Measured with grag-mcp eval-retrieval --rerank against the built-in eval set (13 hand-written cases over the src/graph_rag/eval/corpus/ fixture; 12 of them positive):

The fixture corpus is small and deliberately unambiguous, so every layer already retrieves the right chunk inside the top-k — the differentiator here is how close to rank 1 it lands. Each stage tightens that: full-text catches exact-term matches the bi-encoder blurs, and the cross-encoder promotes a couple of rank-2 hits to rank-1. The layering effect grows as a corpus gets larger and noisier and the top-k stops being trivially correct.

Query rewriting (optional)

Where reranking re-orders a shortlist after retrieval, query rewriting widens what retrieval sees before it. Setting GRAG_QUERY_REWRITE=1 turns one query into a few — expanding acronyms, splitting a multi-part question into sub-queries, paraphrasing toward documentation vocabulary — runs hybrid search for each, and fuses the hit sets (a chunk keeps its best [0, 1] score across the variants) before the reranker/top-k stage. It applies to search, search_code, and search_policies.

Off by default. With the switch on, two backends are available:

• Heuristic (default — no network, no key). A built-in acronym map (k8s → kubernetes, iac → infrastructure as code, …) plus multi-part splitting on and / ; / ,. Extend the map with a JSON file: GRAG_QUERY_REWRITE_SYNONYMS=/path/to/terms.json ({"term": "expansion"}, merged over the built-ins).
• LLM (opt-in). Set GRAG_QUERY_REWRITE_MODEL to a model id and provide a key (GRAG_QUERY_REWRITE_API_KEY, or OPENAI_API_KEY) — the server then fails at startup if no key is set, not at the first query. Calls an OpenAI-compatible POST /v1/chat/completions; point GRAG_QUERY_REWRITE_API_BASE at a local Ollama / LM Studio / vLLM endpoint to keep it on-box. Any failure (network, non-JSON reply, …) silently falls back to the unrewritten query.

GRAG_QUERY_REWRITE_MAX_QUERIES (default 3) caps the original-plus-variants list per search. Measure a pass with grag-mcp eval-retrieval --rewrite (composable with --rerank). As with reranking, the built-in fixture corpus is too small and unambiguous to show a rank delta — every case already retrieves inside the top-k whether rewriting is on or off; the benefit shows on larger, jargon-heavy corpora where a query and the text that answers it don't share vocabulary.

Split deployment (optional)

docker-compose.yml runs the knowledge base and agent memory together — one Neo4j, one MCP server. To deploy them independently instead (separate Neo4j, separate MCP server, no shared process — even separate hosts), use the two opt-in compose files:

docker compose -f docker-compose.knowledge.yml up   # search/ingest/graph tools
docker compose -f docker-compose.memory.yml up      # remember/recall/forget

Both build from the same Dockerfile, targeting knowledge (full parser stack, incl. the [pdf] extra) or memory (embedder + memory module only — no parsers, no pymupdf) — one codebase, two role-appropriate images. An agent connects to both MCP endpoints as two separate servers. about_qualified_name (tagging/filtering a memory by a CodeEntity) works either way — it's a property on AgentMemory, not a graph edge — but the graph-native "what's been remembered about this function" traversal (get_neighbors from a CodeEntity) only works when both share one database, i.e. not in this split.

Each compose file provisions its own fresh Neo4j — run grag-mcp apply-schema against each before first use (see docs/operations.md).

Wiring a coding agent in another project up to remember/recall/forget as its own persistent memory — for Claude Code and VS Code Copilot Chat — is covered in examples/agent-memory/.

Architecture

flowchart TD
    A["Files: PDF / Markdown / Python / Java / JS / TS / SQL / CSS / YAML"] --> B["Ingestion CLI / API"]
    B --> C{"Parser registry (by extension)"}
    C --> C1["PdfParser"]
    C --> C2["MarkdownParser"]
    C --> C3["PythonParser (ast)"]
    C --> C4["JavaParser (tree-sitter)"]
    C --> C5["JavaScriptParser (tree-sitter, JS + TS)"]
    C --> C6["SqlParser (sqlglot: schema + PL/SQL)"]
    C --> C7["StylesheetParser (tree-sitter, CSS/SCSS/Less)"]
    C --> C8["YamlParser (Checkov-aware)"]
    C1 --> D["Structure-aware Chunker"]
    C2 --> D
    C3 --> D
    C4 --> D
    C5 --> D
    C6 --> D
    C7 --> D
    C8 --> D
    D --> E["Enricher (embeddings + optional LLM entity/relation extraction)"]
    E --> F["Graph writer (idempotent upsert by content hash)"]
    F --> G[("Neo4j (Docker)")]
    G <--> H["MCP server (Streamable HTTP)"]
    H <--> I["Coding agent"]

Design and component breakdown: docs/ARCHITECTURE.md. Roadmap and planning: docs/ROADMAP.md. Backup/restore and day-2 ops: docs/operations.md.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. The repo follows a strict one-class-per-file layout; the conventions are spelled out there. By contributing you agree your work is licensed under Apache-2.0.

License

Apache License 2.0. See NOTICE for third-party components.