Amey-Thakur Typo tolerance
Users mistype constantly, and a search that returns nothing for a one-letter error looks broken. Fuzzy matching fixes that and, applied carelessly, makes short and precise queries return nonsense, so the policy has to vary by query.
Method
- Scale tolerance with word length. Short words should tolerate nothing or one edit, longer words two. A two-edit allowance on a four-letter word matches half the dictionary.
- Never fuzzy-match identifiers. Codes, SKUs, and reference numbers must match exactly, since a near-miss result is worse than none (see full-text-search-design).
- Rank exact above corrected. A corrected match should appear below anything that matched as typed, so precision is preserved for users who typed correctly.
- Prefer did-you-mean when confidence is low. Offering a correction keeps the user in control, while silently rewriting hides what happened (see zero-results-handling).
- Use your own corpus to correct, not a general dictionary. The right correction for a domain term is another domain term, which a language dictionary does not contain.
- Weight corrections by popularity. Among plausible corrections, the one people actually search for is usually right (see search-analytics).
- Handle keyboard-adjacent and phonetic errors separately. They have different causes and different fixes, and one setting does not serve both.
Boundaries
- Typo tolerance costs query performance, since fuzzy matching expands the candidate set considerably.
- It cannot distinguish a typo from a deliberately unusual term, which is why exact matches must stay ranked first.
- Correction quality depends on corpus size; a small index cannot correct reliably.