RECALL
初召回重覆盖 Recall favors coverage
在大库里快速找出 k 个可能相关候选。 Quickly find k possibly relevant candidates in a large corpus.
01 · 第一眼相似，不代表最能回答
01 · First-glance similarity is not final relevance
粗召回把“退货概览”排第一，却把真正写明“标准版 30 天”的条款放第四；要不要再仔细读一遍？ Coarse recall ranks a return overview first but puts the clause stating “Standard — 30 days” fourth. Should a stronger judge read them again?
02 双队列：调 k，再开精排 Two Queues: Set k, Then Enable Refinement
候选池 candidate pool
最终第一
Final #1
D4
决定性条款
decisive clause
Top-3 相关
Top-3 relevant
3/3
相关文档
relevant docs
估算延迟
Estimated latency
65
ms
粗召回队列
Coarse recall queue
最终队列
Final queue
03 召回与精排是两种职责 Recall and Reranking Have Different Jobs
快筛 · 细判 fast screen · deep judgePOOL
k 划定可见范围 k sets the visible pool
没进入 Top-k 的文档不会被后续精排重新发现。 A document outside Top-k cannot be rediscovered by the downstream reranker.
RERANK
精排重相关性 Reranking favors relevance
更仔细比较查询与完整候选文本，重新打分。 Compare the query with each full candidate more carefully and rescore it.
LATENCY
精度要付延迟 Precision costs latency
k 越大，精排阅读的候选越多，时间和成本越高。 Larger k makes the reranker read more candidates, increasing time and cost.
04 最小机制：精排只作用于 Top-k Minimal Mechanism: Reranking Acts Only on Top-k
R(Ck)
Ck = top-k(recall(q))
final = sort(rerank(q, Ck))
final = sort(rerank(q, Ck))
若关键 D4 不在 Ck 中，rerank(q,Ck)
无论多强都无法把 D4 排回第一。
If decisive D4 is absent from Ck, no rerank(q,Ck) can place D4
first, however strong the reranker is.
粗召回 Top-k
Recall Top-k
保证候选覆盖
protect candidate coverage
逐候选细读
Judge each candidate
更强相关性分数
stronger relevance score
输出精排顺序
Emit refined order
同时记录延迟
track latency too
05 边界实验：k 太小，关键证据没进门 Boundary Test: k Is Too Small for the Decisive Evidence
主动制造失败 Create a failure把初召回收窄到 k=2 Narrow initial recall to k=2
D4 在粗排第 4，因此根本不会送入精排。精排仍然运行，却只能在错误候选里挑最好。 D4 is fourth in coarse recall and never reaches the reranker. Refinement still runs, but it can only choose the best among the wrong pool.
尚未执行。当前 k=5 包含决定性 D4。
Not run yet. Current k=5 includes decisive D4.
✗ 重排序能找回没召回的文档。
Reranking can recover an unrecalled document.
→ ✓ 它只能重排 Top-k 候选池。
It can only reorder the Top-k candidate pool.
✗ k 越大越好。
A larger k is always better.
→ ✓ 覆盖增加也会带来延迟与成本。
Coverage gains also add latency and cost.
✗ 召回分数就是最终相关性。
Recall score is final relevance.
→ ✓ 粗召回与精排服务不同目标。
Coarse recall and reranking serve different objectives.
06 · 一句话带走
06 · One line to keep