'hybrid': native fusion of dense + lexical (only distinct on backends with native fusion, e.g.
Mongo Atlas; on pgvector it degrades to dense + a legacy app-side keyword fallback).
'keyword': embedding-free lexical retrieval; the mechanism is backend-specific — native Lucene BM25
on Mongo Atlas (ranked, term-optional: a partial-term query still matches), and a boolean substring
filter on pgvector (every term must appear literally; unranked).
'graph': GraphRAG retrieval — seeds entities via vector search, expands relationships via graph
traversal, then maps entities back to their source chunks, RRF-fused with a standard hybrid search.
Only available on vectorDbType:'mongoAtlas' KBs with graphRag.enabled; tuned via
AiKnowledgeBaseGraphSearchOptions. This is the DEFAULT mode when a search omits searchMode
on a graphRag.enabled KB (non-graph KBs default to 'hybrid').
Retrieval mode for a knowledge-base search.
'vector': dense-only similarity.'hybrid': native fusion of dense + lexical (only distinct on backends with native fusion, e.g. Mongo Atlas; on pgvector it degrades to dense + a legacy app-side keyword fallback).'keyword': embedding-free lexical retrieval; the mechanism is backend-specific — native Lucene BM25 on Mongo Atlas (ranked, term-optional: a partial-term query still matches), and a boolean substring filter on pgvector (every term must appear literally; unranked).'graph': GraphRAG retrieval — seeds entities via vector search, expands relationships via graph traversal, then maps entities back to their source chunks, RRF-fused with a standard hybrid search. Only available onvectorDbType:'mongoAtlas'KBs withgraphRag.enabled; tuned via AiKnowledgeBaseGraphSearchOptions. This is the DEFAULT mode when a search omitssearchModeon agraphRag.enabledKB (non-graph KBs default to'hybrid').