Functions, classes, methods and interfaces are embedded and stored in a vector index alongside the graph. An agent asking "where do we verify a webhook signature" gets the functions that do it, ranked by similarity and by how central each one is — not a list of files containing the word "verify". Unchanged symbols are never re-embedded, so re-indexing a quiet repository costs nothing.
- semantic_code_search, selectable on any agent
- 1024-dimension embeddings in an HNSW cosine index
- Vector similarity retrieves; graph structure ranks
- Embeddings cached by content hash and model across commits
- Requires a ready Code Map index for the repository
1024
embedding dimensions