GraphRAG Without the Cloud Bill: A Local-First UI Ecosystem

What it does This project adapts Microsoft’s GraphRAG to run against local models—Ollama for embeddings and LLMs, or any OpenAI-compatible endpoint. It exposes the pipeline through a FastAPI server and provides Gradio-based web interfaces for indexing, prompt tuning, querying, and chatting with your knowledge graph. A legacy main app still handles 2D/3D Plotly visualizations and file management while the architecture transitions toward dedicated single-purpose apps.

The interesting bit The author is mid-refactor, splitting a monolithic Gradio app into an API-centric backend and separate front-ends. It is a rare public admission that the codebase is in a “transition period” with “potential instability,” which feels more honest than most roadmap sections. The whole thing has also been primarily tested on a Mac Studio M2, giving it a distinctly Apple-silicon flavor.

Key highlights

• FastAPI core (api.py) handles indexing, prompt tuning, and global/local/direct chat queries.
• Separate Gradio apps for indexing/prompt tuning (index_app.py) and legacy visualization/querying (app.py).
• Runs fully offline against local models; explicitly aims to eliminate cloud API costs.
• Includes 2D and 3D knowledge graph visualization via Plotly in the legacy UI.
• Bundles its own modified graphrag package with changes not yet in the upstream Microsoft repo.

Caveats

• The ecosystem is undergoing a “major transition”; the author warns users to expect “some changes and potential instability.”
• It has been primarily tested on a Mac Studio M2, so your mileage may vary on other hardware.
• Windows users have reported encoding/UTF-8 errors that need manual YAML adjustments.

Verdict If you want to experiment with knowledge graphs on local hardware and prefer clicking buttons to writing YAML, this is a reasonable starting point—provided you do not mind living through an active refactor. If you need production stability or a fully finished querying UI, check back after the transition.