dsearch2 v7.2 Pipeline
Hybrid semantic disk search. 950K chunks from 128K files. Query flows left→right through alias expansion, 3 active retrieval arms, RRF fusion, then optional AI answer.
Active arms (3): BGE-M3 vector (W=1.0), TF-IDF lexical (W=1.0), Recoll path BM25 (W=1.5). These 3 cover orthogonal axes — semantic, exact-term, filesystem.
Disabled (2): Recoll content (W=0.0, noise) and FTS5 (W=0.0, lexical axis saturated — 0 gains in grid search).
Optional: HyDE via --hyde flag (helps abstract queries, hurts navigational — net −5.7pp, manual only).
Data stores: ChromaDB BGE (vectors), SQLite chunks.db (text + FTS5, 900× faster than ChromaDB text fetch).
38/70 golden hits (54.3%). Easy: 30/38 (79%). Hard: 8/32 (25%). Retrieval-side ceiling reached — further gains need data quality work.
Click a box or arrow for details · drag boxes · bend arrows with yellow dot · positions auto-saved per browser.