retrieval #

query-time mechanics for vector search: how documents and queries get embedded, how keyword and semantic results combine, and what happens between retrieval and the prompt. evidence drawn from pub-search (search over atproto publications) and phi (a bluesky agent's memory reads).

notes #

• asymmetric-embedding — set input_type at both ends, and the document-prep checklist (truncation, titles, utf-8)
• reciprocal-rank-fusion — fusing keyword and semantic results by rank position, and why the two paths must fail independently
• multimodal-documents — one same-sized image per document or it becomes a hub; captions give the text side mass
• synthesize-before-injecting — the stale-memory failure from prompting raw top-k, and the cheap-model pass that fixed it

operational notes on the vector store itself: storage/turbopuffer.

sources #

• pub-search — hybrid keyword+semantic search backend
• bot — phi's namespace memory