Microsoft's GraphRAG, but it runs on your laptop via Ollama

What it does

This is an adaptation of Microsoft’s GraphRAG that routes all LLM and embedding calls to Ollama instead of OpenAI. You feed it documents—.txt, .csv, .json, or .jsonl—and it constructs an entity knowledge graph, then answers questions by traversing communities of related concepts. It runs entirely offline through the OpenAI-compatible endpoint at localhost:11434.

The interesting bit

The project doesn’t just proxy chat completions; it reimplements GraphRAG’s full indexing and query stack for local inference, including a Lazy mode that skips community summarization during indexing and defers it to query time. The README claims that trade-off yields roughly 99% faster indexing, at the cost of slightly slower first queries.

Key highlights

• Five query modes: Global, Local, DRIFT (iterative sub-question decomposition), Basic (vector-only), and Lazy (on-demand summarization).
• Built-in Web UI for indexing, querying, and interactive graph visualization via pyvis.
• Multilingual-aware chunking with UTF-8 enforcement and character-count splitting for CJK, Arabic, and Cyrillic text.
• Docker Compose setup for one-command deployment with GPU passthrough support.
• JSON/JSONL input support alongside plain text and CSV.

Caveats

• The README warns you must stick to Python 3.10 to avoid errors, and explicitly asks for “support for llama integration,” suggesting model coverage is still incomplete.
• DRIFT search works best when community reports are pre-built; using Lazy indexing leaves the primer phase with limited context.

Verdict

Worth a look if you want GraphRAG’s corpus-level sensemaking without cloud dependencies or metered API calls. Skip it if you need polished, production-grade reliability; this is a community adaptation with visible rough edges.