GraphifyHome › Leiden Community Detection
Leiden Community Detection Without Embeddings
Graphify groups related code, docs and diagrams into communities using the Leiden algorithm over graph topology alone. No vector embeddings, no vector database, no separate similarity index.
Why not embeddings?
Most code-RAG systems chunk a repo, embed each chunk, and cluster the resulting vectors. That pipeline has three failure modes Graphify is designed to avoid:
- Lost structure. Embedding similarity ignores call graphs and imports — the exact edges that tell you two files belong together.
- Opaque clusters. When two chunks land in the same bucket you can't point at why. The similarity is a dot-product, not a reason.
- Extra infrastructure. A vector store is another service, another auth boundary, another cost center.
Graphify sidesteps all three by running Leiden directly on the graph the Tree-sitter AST pass and the semantic pass produce. Edge density is the clustering signal.
How semantic similarity still participates
The semantic pass emits semantically_similar_to edges between nodes that look conceptually related but have no structural connection — a function in code and a concept in a paper describing the same algorithm, for example. These edges are marked INFERRED with a confidence score, and they live in the same graph as the structural edges. Leiden sees them, edge density goes up where it should, and communities form around conceptual affinity as well as call structure. No vector index required.
What a community looks like in the output
After clustering, each community becomes a section in GRAPH_REPORT.md and, optionally, a standalone article when you pass --wiki. The report lists:
- the community's top-ranked god nodes — highest-degree concepts inside it;
- the surprising connections — edges into or out of the community that score high on a composite ranker (code-to-paper beats code-to-code);
- a small set of suggested questions this community is uniquely positioned to answer.
For a worked example, the httpx corpus yields 6 communities with god nodes Client, AsyncClient, Response and Request, and surfaces the surprise edge DigestAuth → Response.
Tech choice
Leiden is implemented via graspologic. The rest of the graph layer is NetworkX. The entire clustering stage is pure-Python and runs locally, consistent with Graphify's no-server, no-telemetry posture.