Hindsight swaps MiniLM reranker for Jev: recall@1 rises from 0.80 to 0.95
A retrieval loop uses two Jev calls to score passages and verify claims, at 0.027 s per query with answerable and injection gates.
TypeSafe built Jev to return probabilities, not prose. A typed model that scores a full passage pool in one request and fires two binary gates on the spot. Hindsight swapped their MiniLM reranker for it: 0.027 s vs 0.12 s per query at 30 candidates, recall@1 0.950 vs 0.800 (Hindsight benchmark, LoCoMo). Two Jev calls score and verify the whole loop: 1. HYBRID SEARCH: dense vectors + keyword match surface the top 20 candidates (design value, per the post) 2. JEV #1: one request scores all 20 passages and fires two gates: answerable? injection? 3. THRESHOLD: 0.6 drops everything below it. The writer agent never sees low-confidence material (design value, per the post) 4. WRITER: drafts the answer from the passages that survived 5. JEV #2: walks every claim in the draft back to its source passage 6. GATE FAIL: answerable returns no, writer never fires. A researcher reshapes the query for one retry. Second miss kills the loop: "not in the docs" the injection flag is a retrieval filter, not a security perimeter. remove the answerable gate and every unanswerable query becomes a paid hallucination. two scoring calls, two gates, one hard no. The full loop runs in the terminal below ↓
This post quotes another post: x.com/ridark_eth/status/2112201965123809623