Exact search and semantic search solve different problems
Full-text search is excellent when you remember a customer name, error code, feature label, or exact phrase. Semantic retrieval helps when you remember the meaning but not the wording.
Hybrid retrieval is stronger than either signal alone
A local meeting knowledge base can combine SQLite FTS5 lexical ranking with embedding similarity, then fuse and rerank the candidates before giving evidence to the language model.
Cross-meeting diversity prevents one long call from dominating
When a question spans the entire library, retrieval should not let a single verbose meeting fill every result slot. Diversity across meetings improves the evidence set for longitudinal questions.
Answers should cite the source meeting
A useful cross-meeting answer identifies where the evidence came from: the meeting, transcript location, summary section, or screenshot timestamp.
RAG should retrieve current source artifacts
The index is a derived acceleration layer. At answer time, the system should resolve the latest authoritative transcript, summary, or screenshot evidence instead of treating stale indexed text as the canonical record.