Components

• parsers.py parses raw history files into normalized turns.
• store.py owns SQLite schema, writes, and retrieval.
• embeddings.py provides local hash or OpenAI embedding providers.
• search.py merges lexical + semantic rankings.
• cli.py exposes init, index, search, recall, stats.

Storage model

• prompts table: normalized turns + metadata
• prompts_fts virtual table: FTS5 index for keyword matching
• prompt_embeddings table: serialized vectors
• index_runs table: indexing run audit trail

Retrieval model

1. Lexical candidate set via FTS5
2. Semantic candidate set via sqlite-vec or fallback cosine
3. Weighted hybrid score merge
4. recall format adds next assistant turn as “What worked”