Entity matcher that uses embedding similarity against entity descriptions.
Embeds the query with the configured embedder, then searches the graph store for entities whose description embeddings are most similar via pgvector cosine distance.
This is the default matcher and aligns with Microsoft GraphRAG's Local Search approach. It works best for conceptual or thematic queries where the query describes what you're looking for rather than naming it.
Options
:threshold- minimum cosine similarity (default: 0.3):limit- maximum entities returned (default: 20):repo- Ecto repo (required)
Requires entity embeddings to be populated. Run mix arcana.graph.embed_entities
on existing graphs.