GraphRAG in Rust, now living in your browser

What it does

GraphRAG-rs ingests documents, extracts entities and relationships, and builds a queryable knowledge graph. It ships as a server-backed binary with Qdrant and Ollama integration, a fully client-side WASM build that runs in the browser via WebGPU and WebLLM, and a planned hybrid mode. The library exposes a quick-start API, a typed builder, and a CLI for indexing and natural-language queries.

The interesting bit

The WASM-only deployment is the standout: it performs the entire GraphRAG pipeline—including embedding, vector search, and LLM synthesis—inside the browser using ONNX Runtime Web and pure Rust cosine similarity, with no backend required. The project also bundles five recent research techniques—LightRAG dual-level retrieval, Leiden community detection, cross-encoder reranking, HippoRAG PageRank, and semantic chunking—and claims a 20% accuracy gain and 99% cost reduction.

Key highlights

• Three deployment targets: a 5.2 MB server binary with Qdrant and Ollama support, a fully client-side WASM build with WebGPU acceleration, and a hybrid mode still on the roadmap.
• Browser-native pipeline includes entity extraction, relationship mapping, pure Rust vector search, and WebLLM-powered answer synthesis; the demo indexes Plato’s Symposium into 2,691 entities.
• Bundles research-backed features like LightRAG retrieval, Leiden clustering, cross-encoder reranking, HippoRAG PageRank, and semantic chunking.
• Hash-based fallback embeddings work out of the box without any LLM or vector database, with optional upgrades to Ollama, GLiNER, or custom chunking.
• Compile-time safety via a TypedBuilder API that refuses to build if required fields are missing.

Caveats

• The hybrid deployment mode is architecture-only; the README says implementation is in “Phase 3” and not yet ready.
• Several README links and clone instructions use placeholder URLs (your-username), suggesting the docs are still being polished.
• The advertised accuracy and cost figures are self-reported from the README and lack independent benchmarks in the repo.

Verdict

Worth a look if you need a private, offline knowledge-graph pipeline or a Rust-native GraphRAG alternative. Skip it if you require a mature hybrid client/server sync or are skeptical of self-reported accuracy leaps.