Graph your documents with a local LLM and zero cloud spend

What it does Splits a body of text into chunks, then uses a local Mistral 7B model to extract concepts and their relationships. It treats co-occurrence within a chunk as a contextual edge, merges duplicate pairs, and hands the result to NetworkX and Pyvis for visualization and basic graph analytics like degree and community detection. Everything stays on your machine; there are no calls to OpenAI.

The interesting bit The author deliberately extracts concepts rather than raw entities—so “Pleasant weather in Bangalore” becomes a node, not just “Bangalore”—and weights edges by both LLM-inferred relationships and simple contextual proximity. The result is a graph schema built in Pandas, which is either admirably pragmatic or a reminder that this is still a notebook-sized experiment.

Key highlights

• Runs fully offline using Mistral 7B via Ollama; no API keys or cloud spend.
• Extracts higher-level concepts instead of just named entities, aiming for richer semantic nodes.
• Collapses multiple relationships between the same concept pair into a single weighted edge.
• Uses Pyvis to generate interactive web-ready visualizations (see the repo’s GitHub Pages link).
• Ships as a Jupyter notebook (extract_graph.ipynb) rather than a packaged library.

Caveats

• The author flags that deduplication is still manual: semantically identical concepts like doctor and doctors can spawn separate nodes until embeddings are added.
• There is no dedicated frontend for exploration yet; interaction is limited to the Pyvis visualization or notebook output.
• Graph storage is a set of Pandas dataframes, so this is an analytics script rather than a persistent knowledge base.

Verdict A useful sandbox if you are experimenting with Graph RAG and need a fully local, zero-cost pipeline to turn documents into concept webs. Skip it if you need a hardened, deduplicated graph database backend or a polished query interface out of the box.