LLMs, GenAI, RAG, agents, evaluation · insight-1535
Extract metadata filters from the query with a schema, and fall back to unfiltered search when they return nothing
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Applied AI, ML engineer
- Industries
- aviation, e-commerce
Summary: situation, action, reason, boundary
Situation
Users type "IST-FRA business fares after March" and the index has route, cabin and effective-date fields that semantic search ignores.
Action
Parse the query into a typed filter object, validate values against the known vocabularies, apply them as hard filters, and when results fall below a minimum retry unfiltered while telling the generator which filter was dropped.
Reason
Filters turn entity mentions into exact constraints, but extraction itself is a fallible model step, so a failed or over-narrow filter must degrade gracefully.
Boundary
Catalogues, policy sets and ticket archives with reliable structured fields.
Symptom or risk
Unfiltered retrieval returns semantically close chunks from other routes, while a wrong extracted filter returns nothing and the assistant answers from thin air or refuses.
Exceptions
Queries where the entity is the answer, such as "which routes allow free changes".
Trade-offs
An extra parse step adds latency and a failure mode that needs its own tests.
Pitfalls
Silently dropping a filter, so the answer quotes a rule for the wrong route without any signal.