Add FTS5 hybrid search (keyword + semantic)

[jonyoder]

The sqlite-vec database already uses SQLite. Adding an FTS5 table alongside the vector table would enable hybrid search — keyword matching for exact function/type names combined with semantic vector search for conceptual queries.

The code_items metadata table is already there; just need to create an FTS5 virtual table indexing the same data.

This would improve results for queries like exact function names (where keyword match is better) while keeping semantic search for conceptual queries.

[m13v]

totally agree on the hybrid approach. we built exactly this, FTS5 table alongside sqlite-vec for the same dataset. for function and type name lookups, FTS5 with BM25 scoring wins every time since embeddings don't preserve exact identifiers well. for conceptual queries like 'error handling pattern' the vector search shines. the ranking blend we settled on was query-length dependent: short queries (under 5 tokens) lean 70/30 toward keyword, longer natural language questions go 30/70 toward semantic. keeps the best of both worlds without the user needing to know which search mode to use.

[m13v]

our hybrid FTS5 + sqlite-vec implementation with query-length-dependent scoring blend: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/embeddings.py

[jonyoder]

@m13v Thanks for the feedback and link. We'll keep this in mind!