DENSE
稠密路找近义 Dense finds paraphrases
D2“常见故障排查”没有查询原词，却语义最接近。 D2 “common troubleshooting” lacks the query wording but is semantically closest.
01 · 一个查询有两种“相关”
01 · One query, two kinds of relevance
“ZX-42 设备出问题”既要找故障语义，也不能漏掉精确编号；两路结果谁先谁后？ “ZX-42 device acting up” needs troubleshooting semantics without losing the exact ID. How should the two result lists combine?
02 融合调音台：给两路检索分配话语权 Fusion Mixer: Give Each Retriever a Voice
融合权重 fusion weight
语义覆盖
Semantic coverage
✓
D2 · troubleshooting
精确覆盖
Exact coverage
✓
D1 · ZX-42
Top-3 覆盖
Top-3 coverage
2/2
目标意图
target intents
03 融合不是把列表拼接起来 Fusion Is Not List Concatenation
归一化 · 合并 · 去重 normalize · merge · dedupeSPARSE
稀疏路抓编号 Sparse catches the ID
D1 因精确包含 ZX-42 获得高词项分数。 D1 earns a strong lexical score by containing ZX-42 exactly.
NORMALIZE
不同分数先对齐 Align incompatible scores
相似度与 BM25 尺度不同，不能直接裸相加。 Similarity and BM25 use different scales and cannot be added raw.
FUSE
合并后去重排序 Merge, dedupe, rerank
同一文档可能两路都出现，最终只保留一项综合排名。 A document can appear in both paths but receives one fused rank.
04 最小机制：融合分数或融合名次 Minimal Mechanism: Fuse Scores or Fuse Ranks
αD + (1−α)Sweighted = αD + (1−α)S
RRF = Σ 1 / (k + rank)
RRF = Σ 1 / (k + rank)
加权融合依赖可比分数；RRF 只使用各路名次。切换规则时，即使权重不变，排序也可能改变。
Weighted fusion needs comparable scores; RRF uses ranks only. Switching rules can change the order even when the balance stays fixed.
两路召回
Two recalls
语义 + 词项
semantic + lexical
按规则融合
Apply fusion rule
分数或名次
scores or ranks
统一候选表
Unified candidates
同时覆盖两种意图
cover both intents
05 边界实验：把稀疏路权重归零 Boundary Test: Zero Out the Sparse Path
主动制造失败 Create a failure只保留稠密检索 Keep dense retrieval only
语义故障文档仍会靠前，但精确含 ZX-42 的 D1 可能掉出 Top-3。两路都“开启”也不够，权重为零等于没有话语权。 Semantic troubleshooting remains high, but D1 containing exact ZX-42 can drop from the Top 3. A path with zero weight has no voice even if technically enabled.
尚未执行。当前 Top-3 同时覆盖语义与编号。
Not run yet. The current Top 3 covers both semantics and the identifier.
✗ 混合检索就是两路都开。
Hybrid retrieval just turns both paths on.
→ ✓ 还要定义归一化、权重与融合规则。
It also needs normalization, weighting, and a fusion rule.
✗ 两个原始分数可直接相加。
Two raw scores can be added directly.
→ ✓ 不同检索器的分数尺度可能完全不同。
Retriever score scales may be entirely different.
✗ 50/50 一定最公平。
50/50 is always fairest.
→ ✓ 配比要用代表性查询集评测。
Tune the balance on a representative query set.
06 · 一句话带走
06 · One line to keep