83k→3.7k compression

It indexes a codebase's domain terminology and behavioral similarity separately, enabling it to answer questions one level above search.

• Questions like "what does dependency mean in this codebase" are handled at the concept level, not string search.
• Functions that "behave similarly to authenticate()" are found via embedding similarity.
• It also catches naming consistency, terminology drift, and duplicate validation logic.

The author states they tested it on FastAPI, PyTorch, voxelmorph, and ScribblePrompt, and that it supports Python/TS/JS/Rust.

The implementation combines tree-sitter + TF-IDF + 2 embedding models + PageRank, and runs entirely locally, requiring no API key.

In a Claude Code-based example, the same task that used to take 27 tool calls, 83k tokens, and 3 minutes reportedly dropped to 4 calls, 3.7k tokens, and 5 seconds.