What is Relevance?
Relevance is the degree of usefulness of a result for a given query and context. It covers topicality (does it match?), utility (is it in stock, the right size/price?), and trust (ratings/returns), all tempered by the user’s intent (navigational, informational, transactional).
How It Works (quick)
- Evidence: Lexical signals (BM25, phrases/proximity), semantic similarity (vectors), entity/attribute matches.
- Context: User locale/device, category, personalization (opt-in), and business constraints (ACL, stock).
- Scoring: Combine evidence into a relevance score; optionally LTR for ordering.
- Evaluation: Measure Precision/NDCG/Recall, zero-results, reformulations, and satisfaction.
Why It Matters in E-commerce
- Drives CTR, conversion, and margin by showing the most suitable SKUs first.
- Reduces waste clicks and facet thrashing; improves perceived quality.
Best Practices
- Keep exact/phrase logic for brands/SKUs; add vectors for meaning.
- Encode must-have business rules early (stock, region, compliance).
- Use category/localespecific weights or models; footwear ≠ laptops.
- Cap popularity/freshness so they don’t outrank clear matches.
- Maintain a golden set and run regular A/B tests.
Challenges
- Noisy labels (position bias), seasonality, catalog churn, multilingual analyzers.
Examples
- “merino base layer” → prioritize in-stock merino tops with strong phrase hits and good return rates.
- “gift card” → route to the brand’s landing page (navigational intent).
Summary
Relevance is matching intent + quality. Use hybrid evidence, enforce hard rules, localize tuning, and validate with golden sets and A/B tests.
