From 4865f974b671d72a1537af918e0bf01da67a14f5 Mon Sep 17 00:00:00 2001
From: Luiz Gustavo <gutomec@gmail.com>
Date: Thu, 21 May 2026 14:46:49 -0300
Subject: [PATCH] fix search: rerank-gate results so absent terms return
 nothing
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The hybrid_search RPC always returns up to recall_k dense neighbours, so a
query for a term absent from the corpus (e.g. "varginha") returned its 12
nearest vectors — irrelevant chunks like PAGE_NUMBER "1". Two bugs:
the reranker was skipped whenever results <= top_k, and there was no relevance
floor.

Now always run the cross-encoder reranker (BGE-reranker-v2-m3, normalized
sigmoid) and drop hits below 0.02. Verified: "varginha" → 0 results;
"roswell"/"tic tac"/"disco voador" → relevant hits on top (reranker cleanly
separates 0.0001 garbage from 0.03-0.27 matches).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 web/lib/retrieval/hybrid.ts | 27 +++++++++++++++++----------
 1 file changed, 17 insertions(+), 10 deletions(-)

diff --git a/web/lib/retrieval/hybrid.ts b/web/lib/retrieval/hybrid.ts
index b157868..df16bbb 100644
--- a/web/lib/retrieval/hybrid.ts
+++ b/web/lib/retrieval/hybrid.ts
@@ -85,25 +85,32 @@ export async function hybridSearch(opts: HybridSearchOptions): Promise<ChunkHit[
 
   if (rows.length === 0) return [];
 
-  // 3. Optional rerank
-  if (no_rerank || rows.length <= top_k) {
+  // 3. Rerank — the cross-encoder score is BOTH the relevance ordering AND the
+  // noise filter. The RPC always returns up to recall_k dense neighbours even
+  // when the term has zero lexical match (e.g. a name absent from the corpus),
+  // so without this filter a nonsense query returns its k nearest vectors.
+  // BGE-reranker-v2-m3 with normalize=true → sigmoid 0..1; ~0 means irrelevant.
+  const RELEVANCE_MIN = 0.02;
+
+  if (no_rerank) {
     return rows.slice(0, top_k);
   }
 
-  const candidateTexts = rows.map((r) => {
-    if (lang === "en") return r.content_en || r.content_pt || "";
-    return r.content_pt || r.content_en || "";
-  });
+  const candidateTexts = rows.map((r) =>
+    lang === "en" ? r.content_en || r.content_pt || "" : r.content_pt || r.content_en || "",
+  );
   let scores: number[] = [];
   try {
     scores = await rerank(query, candidateTexts);
   } catch {
-    // Reranker unavailable — return RRF order
+    // Reranker unavailable — fall back to RRF order (no relevance filter possible)
     return rows.slice(0, top_k);
   }
-  const reranked = rows.map((r, i) => ({ ...r, rerank_score: scores[i] }));
-  reranked.sort((a, b) => (b.rerank_score ?? 0) - (a.rerank_score ?? 0));
-  return reranked.slice(0, top_k);
+  return rows
+    .map((r, i) => ({ ...r, rerank_score: scores[i] ?? 0 }))
+    .filter((r) => (r.rerank_score ?? 0) >= RELEVANCE_MIN)
+    .sort((a, b) => (b.rerank_score ?? 0) - (a.rerank_score ?? 0))
+    .slice(0, top_k);
 }
 
 /** Quick chunk lookup by chunk_id (no embedding). For citation expansion. */