Fuzzy Search

What is Fuzzy Search?

Fuzzy search matches near-by strings rather than exact text, typically using edit distance (e.g., Levenshtein) or phonetic similarity. It tolerates insertions, deletions, substitutions, and transpositions so misspellings and spacing issues still retrieve results.

How It Works (quick)

• Edit distance: Allow 1–2 edits depending on term length.
• Token strategy: Apply fuzziness to unigrams; keep bigrams/phrases stricter.
• Thresholds: Length-aware rules (e.g., no fuzz for 1–2-char tokens; 1 edit for 3–5; 2 edits for ≥6).
• Costs & boosts: Weight edits (transposition cheaper), penalize low-quality matches.
• Fallbacks: Use exact/phrase boosts first; relax progressively when recall is low.

Why It Matters in E-commerce

• Real-world typos: Brand names, model numbers, accents/diacritics.
• Mobile input: Fat-finger errors, missing spaces (“airmax”).
• Locale variants: “Sneakers” vs “sneekers”; “gore tex” vs “gore-tex”.

Best Practices

• Guardrails: Don’t apply fuzz to SKU/MPN fields; keep an exact field.
• Length-aware fuzz: Scale edits with term length; disable for stopwords.
• Synonyms + normalization: Diacritic/case folding; synonym maps for common variants (gtx ↔ GORE-TEX).
• Precision first: Exact/phrase/bigram scores outrank fuzzy hits; cap fuzzy impact.
• Analytics: Monitor “fuzzy saves” vs false positives; tune per locale.

Challenges

• Over-recall: Irrelevant results from aggressive fuzziness.
• Performance: Fuzzy matching is costlier; cache frequent queries.
• Brand integrity: Don’t degrade protected brand casing/spacing.

Examples

• “nik” → Nike (1 edit); “gore tex” → GORE-TEX (space + case).
• “iphon 15” → iPhone 15 (substitution).
• “airmax” → Air Max (missing space).

Summary

Fuzzy search rescues typo-ridden queries by allowing small edit distances—without sacrificing precision when you cap its influence, protect exact fields, and combine with synonyms and phrase/bigram logic.