What is Recall?
Recall is an IR metric:
recall = relevant_results_returned / total_relevant_results.
High recall means you retrieved most of what mattered, even if some noise slips in.
How It Works (quick)
- Evaluate: Labeled sets or debiased click data → compute Recall@k and coverage.
- Boost recall: Synonyms (late-bound), fuzzy/typo tolerance, phrase relax, vector/dense recall, category expansions.
- Guardrails: Keep exact SKU/brand strict; cap expansions; enforce ACL/stock.
Why It Matters in E-commerce
- Fewer dead ends: Prevents zero-results and “no good options” experiences.
- Long-tail wins: Captures niche queries with specific intent.
- Resilience: Handles wording mismatches between users and catalog.
Best Practices
- Run hybrid retrieval (BM25 + vectors) with late-bound synonyms.
- Apply length-aware fuzziness (off on SKUs/brands).
- Phrase relax if exact phrase fails; widen proximity narrowly.
- Track zero-results rate, Recall@k, reformulations, and assisted revenue.
- Segment by category/locale; tune expansions where gaps are largest.
Challenges
- Over-expansion harms precision; multilingual ambiguity; compute cost for vectors; evaluation noise for rare queries.
Examples
- “gtx trail runners” → expand GTX ↔ GORE-TEX, allow small proximity, add vector recall to surface waterproof shoes.
- Cold-start items gain visibility via semantic recall before behavior accumulates.
Summary
Recall is coverage. Use hybrid retrieval, careful fuzziness, and category-aware synonyms to find more good items while protecting precision with strict rules and caps.
