OutDept
Case · RAG and quality

Hybrid search and a self-correcting layer

A knowledge-base retrieval built as a hybrid of vector and full-text search, plus a deterministic guard layer around the model that turns every failure into a permanent rule.

Hybrid retrieval core: vector and FTS candidates fused by reciprocal rank
Hybrid retrieval core: vector and FTS candidates fused by reciprocal rank

What we built

  • Hybrid retrieval: embeddings + SQLite FTS5, merged through reciprocal rank fusion
  • 14,482 chunks indexed — ingestion filtered out 96% of the noise
  • 25 deterministic hooks before and after each answer: fact checks, no claim without a source

Tech stack

Python · fastembed · sqlite-vec · FTS5 · RRF fusion

Got a project like one of these?

Tell us the problem, not the service name — we'll scope it properly before quoting anything.

Talk to us

AI platforms

← All case studies