How can I make my structured data search robust to different query phrasings without duplicating my index?

The position and phrasing of constraints significantly affects retrieval quality in token-based systems. ColBERT showed accuracy drops when "for 5 guests" moved from beginning to end of the query because it relies on token-position interactions. The solution is using an LLM as a query preprocessor that extracts intent and parameters regardless of phrasing. Configure it to identify:

1. hard filters that must be enforced (guest capacity ≥ 5),
2. soft preferences with directional weights (lower price = negative weight), and
3. semantic concepts that need interpretation ("cozy" → certain amenities/descriptions).

This approach maintained 100% constraint satisfaction across query variations in the benchmark, while ColBERT's accuracy varied by 40% based on phrasing.

Pro tip: Cache the LLM's parameter extraction for common query patterns to reduce latency.