Failure Modes and Production Guardrails

Over-Correction and User Frustration

Aggressive spell correction changes valid queries. "Golang" corrects to "goal" in a general-purpose search. Brand names, technical terms, and proper nouns are especially vulnerable. Users who repeatedly see wrong corrections lose trust. Prevention: maintain allowlists of valid terms by domain, require higher confidence for corrections on capitalized words, and always provide "search instead for [original]" option.

Expansion Drift

Query expansion adds semantically similar but contextually wrong terms. "Python book" expands to include "snake" in a programming context. "Apple stock" expands to "fruit" when user wants financial data. Drift compounds: each expansion term can trigger further expansions. Prevention: use entity linking to constrain expansions ("Python" linked to programming language excludes animal expansions), limit expansion depth, validate expansions against user click patterns.

Intent Misclassification Cascades

Wrong intent routes to wrong backend, returning completely irrelevant results. A transactional query hitting an informational index returns articles instead of products. User abandons search. Unlike ranking errors (wrong order of good results), intent errors return the wrong result set entirely. Mitigation: use confidence thresholds for routing; if intent classifier is uncertain (<70% confidence), query multiple backends and blend results.