Docs → Memsense Docs
See also: Architecture Overview · Retrieval Algorithm
This page gives a compact summary of how Memsense stores embeddings and performs search.
- save flow writes cleaned QA chunk metadata into
memory_chunks - structured metadata noise is stripped before ingest
- worker computes full/user/assistant/next-user/facet embeddings and stores vectors in
memory_chunk_embeddings - search uses 8-route recall, SQL RRF fusion, session-first hybrid selection when session chunks are present, and MMR diversity selection
- OpenAI-compatible:
MEMSENSE_EMBEDDING_PROVIDER=openai - Local BGE:
MEMSENSE_EMBEDDING_PROVIDER=bge_http
Search returns ranked chunks with fields such as:
final_scorerrf_scoreroutesexplain
- Read Retrieval Algorithm for scoring details.
- Read Architecture Overview for the full system flow.