Single-modality retrieval 會喺 edge cases 失手：dense vectors 漏 rare tokens 同 IDs；pure lexical 漏 paraphrase 同 semantic similarity。Hybrid retrieval 將互補 signals 融合 — dense semantic、sparse lexical、structured metadata、temporal freshness — 產生穩定 high-precision candidate sets。呢篇文章講 architecture、normalization、scoring fusion、failure handling 同 evaluation。
Motivation
Failure scenarios：
- Dense model 漏 proper nouns / SKU codes。
- Pricing change queries 拉到 stale snapshot，因為冇 temporal boost。
- 長 natural questions 喺 sparse-only system 入面被 stopwords overweight。
- Semantically broad pages，例如 marketing fluff，產生 vector false positives，但冇 lexical anchoring。
Hybrid 透過捕捉互相獨立嘅 evidence dimensions 去 mitigate。
Component Layering
Recommended flow：
- Query Embedding -> ANN search (k_vec)
- Lexical Search (BM25 / SPLADE / Elasticsearch) (k_lex)
- Union -> Score Normalization (per source scaling)
- Metadata Filter Pass (locale, access_tier, page_type)
- Diversity & Freshness Adjustments
- Optional Cross/Mono Re-Ranker
- Final Truncation (top K)
保留 raw pre-fusion scores 供 audit。
Query Normalization
Steps：
- Unicode normalize NFKC
- Lowercase，但如有需要，保留 casing snapshot 供 answer formatting
- Tokenize 並 preserve stopwords，因為 semantic embeddings 可以利用 context
- Synonym / Alias Expansion：append alternative tokens，用於 internal product codename mapping；唔插入 model prompt，只用於 sparse retrieval。
- Numeric & Version Extraction：捕捉 X.Y.Z patterns 做 targeted lexical scoring。
Metadata & Attribute Filters
喺 initial candidate union 之後套用 filters，可以減少 recall loss。Common fields：locale、access_tier、page_type、product_area、updated_bucket。Security filters (tenant / tier) 要喺 scoring fusion 之前 enforce，避免 leakage 影響 re-ranking。提供 debug mode 回傳 filtered_out set 供 inspection。
Re-Ranking Strategy
喺 top N (10-20) 上用 lightweight cross-encoder，例如 distilled model。如果 latency > budget，就 degrade：skip re-rank，或者降低 candidate count 同時提高 lexical weight。追蹤 re_rank_delta = MRR_post - MRR_pre 去 justify cost。對 short TTL 內 identical union sets cache re-rank results。
Freshness & Temporal Signals
計 freshness_weight = exp(-lambda * age_days)，lambda 按 content type tune，例如 pricing 高啲、API stable 低啲。合併：final_score = w_sem * sem_score + w_lex * lex_score + w_fresh * freshness_weight + w_meta * meta_priors。先 normalize 每個 component，z-score 或 min-max，避免 dominance。
Failure Modes
| Failure | Cause | Mitigation |
|---|---|---|
| Popularity Bias | Overweight lexical tf-idf | Cap term frequency contribution |
| Stale Results | Freshness weight mis-tuned | Recalibrate lambda using evaluation set |
| Locale Leakage | Late filter application | Move security filters earlier |
| Semantic Drift | Embedding model upgrade | Dual-index and A/B compare before rollout |
| Over-fusion Noise | Unbounded union size | Limit union, diversity pruning |
Evaluation Framework
Experiments：
- Ablation：(vector only、lexical only、hybrid w/o rerank、full) 量度 Recall@k、MRR。
- Fusion Weight Tuning：用 validation gold set grid search weights。
- Latency Budget：追蹤每個 configuration 嘅 mean + P95 retrieval latency。
- Drift：每週 monitor head vs tail queries 嘅 recall relative change。
保留帶 config hashes 嘅 evaluation manifest。
Optimization Loop
Cycle：
- Log retrieval traces：query、candidates、scores、source_tag。
- 找出 mis-hits：downstream faithfulness 低或 citation count 低 -> classify root cause，例如 missing lexical candidate、semantic false positive、stale content。
- 調整 weights / thresholds；跑 offline suite。
- 將新 fusion weights 放喺 feature flag 後面 canary。
- 如果 improvement statistically significant，就 promote。
重點
- Hybrid retrieval 係一套可 tune 嘅 dials，要不斷 instrument。
- Security & access filters 要早套用；避免 leakage 入 scoring。
- Re-ranking 必須用可量度 MRR / Recall lift justify latency。
- Temporal decay 防止 outdated、high-authority pages 主導結果。
- Fusion changes 當 code 處理：version、evaluate、roll forward 或 back。