Semantic-only search keeps missing queries that cite exact ordinance numbers like “12.04.210.” What change should the RAG pipeline make?
Select an answer to reveal the explanation.
Short Explanation
Looking for “12.04.210” with vibes-only search is like finding a house by describing the paint color and ignoring the street number. Add keyword or exact-match filters so ordinance IDs actually resolve.
Full Explanation
Semantic-only search often misses queries that cite exact ordinance numbers such as “12.04.210” because embeddings optimize topical similarity, not brittle alphanumeric identity. The appropriate RAG change is to add keyword or exact-match filters alongside semantic retrieval so citation-style queries hit the right section while paraphrase queries still benefit from vectors.
Removing all embeddings and relying on a random document sampler destroys retrieval quality. Lowering generator temperature without changing retrieval only affects fluency of whatever wrong chunk arrived—it does not fix miss rates on exact citations. Truncating ordinance numbers from queries before embedding throws away the strongest exact-match signal residents provide.
Exam caveat: exact-match filters require clean section identifiers in the index metadata; messy OCR of numbers can still fail. Operational check: run a citation-heavy query suite with and without keyword/exact filters and confirm recall on numbered ordinances rises without collapsing semantic performance on paraphrase-only questions. Index ordinance numbers as first-class metadata fields to make exact filters reliable under load. Regression-test a fixed list of numbered ordinances after each index rebuild to catch metadata regressions that break exact-match filters.