The query is embedded with the collection's pinned embedding profile and fused with BM25, section lookup, and multi-size chunks.
Semantic retrieval
Requirements
Indexed vectors, a matching collection embedding profile (text-embedding-3-large at 3,072 dimensions, gemini-embedding-001 at 3,072 dimensions, google/embeddinggemma-300m at 768 dimensions, or voyage-context-4 at 1,024 dimensions), and agent:run. For collections already indexed with OpenAI text-embedding-3-large, query embedding can fall back to the deployment OpenAI key when no organization OpenAI key is connected.
In collection retrieval (Hybrid Retrieval v2), dense pgvector search runs alongside Hebrew/English BM25 keyword search, section-map lookup, vocabulary rules on discovered collections, and optional extra chunk sizes (500 / 1,500 / 4,500), ranked by the workspace’s active ranking profile in Context Engine. Coverage in the response can be complete, partial, or unavailable when every searched collection failed. Partial coverage is not a full-corpus claim.