Architecture
SenterPoint stores knowledge as atoms — one precedent each: a fix, a decision, a gotcha, with citations back to the commits and files it came from. Everything below is about recalling the right atoms fast, at the right price.
Auto-inject hits A only — sub-300 ms, effectively free. Explicit queries run A + B in parallel, fuse via reciprocal-rank fusion, and hand the shortlist to C. The cheap path runs on every session. The expensive one is opt-in.
Full-text search + structured filter (site / tags / kind).
Runs on every session start to auto-inject the most relevant atoms into the model context — sub-300 ms end to end. No LLM call, no embedding round-trip.
Cosine similarity over 768-d local embeddings.
Runs independently of A, not as a refilter. Both paths return their top candidates; the fusion strip below combines them.