graphrag-rs

A Rust implementation of GraphRAG (by automataIA, open-source): it builds a knowledge-graph from documents and answers natural-language queries over it. Where sift-kg stops at extracting the graph, graphrag-rs adds the half sift-kg lacks — the graph-based retrieval layer — so it spans the whole loop from documents to answers.

What it does (the full GraphRAG loop)

Extract a knowledge graph — entities + typed relationships pulled from text (the same document→graph step as sift-kg).
Community detection — runs the Leiden algorithm to cluster the graph into communities, then PageRank to rank within them. This is the move that defines GraphRAG and that the wiki's other KG tools don't make.
Retrieve + generate — graph-based semantic search feeds an LLM that synthesizes the answer. So the extracted graph is turned back into answers, not just browsed.

Why it matters here — the global-query gap

The wiki's retrieval-augmented-generation page tracks a taxonomy of what plain vector RAG misses: no-accumulation, exact-token, factual-connection, temporal-validity. graphrag-rs brings the GraphRAG answer to a fifth: global / corpus-sensemaking queries ("what are the main themes across the whole corpus?"). Community detection + community summaries let the system answer holistic questions over an entire document set that chunk-level vector retrieval structurally can't — the case graphrag exists for. It complements the factual-connection lift the typed knowledge-graph already gave (gbrain's +31.4 P@5), at a coarser, whole-corpus grain.

The distinctive bit — Rust + local-first, everywhere

A trait-based portable core (graphrag-core) ships four ways: a CLI/TUI (graphrag-cli), a REST server with Qdrant vectors (graphrag-server), and a browser-native WASM build with WebGPU acceleration (graphrag-wasm). Async via Tokio; embeddings via ONNX Runtime (GPU); local models via Ollama. So it's a local-first / on-device take on GraphRAG — runnable in the browser with no server — distinct from the cloud-LLM pipelines the rest of this cluster assumes.

Tier note (T3)

Project README / self-description of an open-source tool — concrete (named crates, real algorithms, deployment targets) but vendor-authored, not independently evaluated. Maturity is self-reported ("production-ready Phase 1; 214+ tests, 100% pass; Phase 2 WASM/UI ~60%"); extraction/retrieval quality is asserted, not benchmarked against alternatives.

Re-verified 2026-08-30 (quality cycle) — and it has not moved

Stale at 62 days; re-read against the repository. 525★ / 49 forks, 0 open issues, MIT, not archived — and the last push is 2026-06-02 (GitHub API, 2026-08-30). That date is before this page was written on 2026-06-29, so the project had already stopped when the corpus ingested it, and nothing has happened in the three months since.

That matters for how the self-reported maturity above should be read. "Production-ready Phase 1; 214+ tests, 100% pass; Phase 2 WASM/UI ~60%" is a snapshot of an unfinished roadmap, and Phase 2 is where it stopped. Zero open issues on a dormant repository is not a health signal either — it is what an untended tracker looks like when nobody is filing.

graphrag · knowledge-graph · retrieval-augmented-generation · sift-kg · gbrain · hybrid-retrieval-rag · tools-for-thought