From 085a407f68022367fa167863c1d2e4c858bb38aa Mon Sep 17 00:00:00 2001 From: xiaoyi Date: Mon, 16 Nov 2026 20:43:01 +0800 Subject: [PATCH] =?UTF-8?q?feat(recall):=20ZHIYI=5FRERANK=5FHOLD=EF=BC=88r?= =?UTF-8?q?erank=20=E4=BF=9D=E4=BD=8D=EF=BC=8C=E9=BB=98=E8=AE=A4=E5=85=B3?= =?UTF-8?q?=EF=BC=89=E2=80=94=E2=80=94=E5=AE=9E=E6=B5=8B=E4=BB=BB=E4=BD=95?= =?UTF-8?q?=20N=20=E9=83=BD=E8=B4=9F=E6=94=B6=E7=9B=8A=EF=BC=8C=E7=95=99?= =?UTF-8?q?=E6=A1=A3=E4=B8=8D=E5=86=8D=E9=87=8D=E8=AF=95?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 主尺 84 DA 题同窗口配对(no_cache,漂移 0.0000): hold=0 0.5258 基准 hold=1 0.4881 -3.8pp hold=2 0.3591 -16.7pp(24 题掉零) hold=3 0.3532 -17.3pp(26 题掉零) 原因:向量序头部正是 cross-encoder 要压下去的空壳句,钉住它=废掉 rerank 的主要作用。 11-12 离线重放「保 top3 最好」的结论不成立——那是在向量 cosine 代理尺上测的,与真 cross-encoder 行为相反。 默认 0 ⇒ MMRSelect 原样调用、逐位等价;代码保留为常开旋钮(观测/回归用),生产不设此 env。 --- go/internal/storage/recall.go | 30 +++++++++++++++++++++++++++++- 1 file changed, 29 insertions(+), 1 deletion(-) diff --git a/go/internal/storage/recall.go b/go/internal/storage/recall.go index 471dd61..ea390ce 100644 --- a/go/internal/storage/recall.go +++ b/go/internal/storage/recall.go @@ -1138,7 +1138,35 @@ func (p *RecallPipeline) recallUncached(ctx context.Context, query, namespace st return nil, RecallMeta{CandidatesScanned: poolSize, Truncated: truncated, Status: "partial", PartialReason: "deadline_exceeded"}, nil } stMMR := time.Now() - finalIndices := MMRSelect(reranked, candidates, topK*2, diversity) + // 2026-11-16 t_rerank_hold:rerank「保位」——把**重排前向量序**的前 N 条钉死在头部, + // 只让其余条目参与 cross-encoder/MMR 竞争。 + // 背景:真 rerank 上线后主尺 0.2639→0.5258(+99%),但 6 道 n_da=1 的窄 how-to 题 + // 从 1.00 掉到 0(cross-encoder 把唯一答案那一批压出了 top5)。11-12 离线重放同结论: + // 全重排 0.3724 < 保 top3 0.3900。默认 0 ⇒ MMRSelect 原样调用,逐位等价现状。 + hold := envInt("ZHIYI_RERANK_HOLD", 0) + var finalIndices []int + if hold > 0 && hold < len(candidates) { + keep := hold + if maxKeep := topK * 2; keep > maxKeep { + keep = maxKeep + } + finalIndices = make([]int, 0, topK*2) + for i := 0; i < keep; i++ { + finalIndices = append(finalIndices, i) + } + rest := make([]models.RerankResult, 0, len(reranked)) + for _, r := range reranked { + if r.Index >= keep { + rest = append(rest, r) + } + } + if need := topK*2 - len(finalIndices); need > 0 { + finalIndices = append(finalIndices, MMRSelect(rest, candidates, need, diversity)...) + } + log.Printf("[recall][hold] rerank 保位前 %d 条(共 %d 候选,池 %d)", keep, len(candidates), len(finalIndices)) + } else { + finalIndices = MMRSelect(reranked, candidates, topK*2, diversity) + } // RRF 开启时多取一倍候选进融合池:只在 topK 内部重排不会改变 top5 的**集合**(等于没效果), // 第 6~10 名必须有机会被词面秩抬进前 5,融合才可能有增益。门控关 ⇒ poolN==topK,逐位等价现状。