Lesson 09: Keyword and semantic search

One renewal rule can be queried as renewal duration or pushing a return date back. The first shares source terms; the second may not. Retrieval methods can respond differently.

Keywords retain explicit terms

Keyword search examines term occurrences. Real systems tokenize, process text, and rank results rather than require identical sentences. Original titles, identifiers, and dates can be valuable signals.

For opening on 18 October 2026, retain the date. A generic opening query can retrieve A and miss D. Exact query details also help judge applicability.

Semantic search handles different wording

Semantic search compares meaning, often through vectors. Pushing a date back can connect to renewal without requiring exact source wording. Original and extended loan periods remain similar enough to require answerability checks.

Keywords can miss paraphrases; semantic retrieval can match the topic with wrong conditions. Test actual queries. E is useful for historical limits and potentially misleading for current ones.

Rewrite without changing the question

A return-date extension can be rewritten as renewal duration. Rewriting it as original loan period changes intent. Recheck object, time, negation, and conditions rather than maximize matches.

Some queries need exact dates and paraphrase understanding together. A later article combines their candidates. Observe them separately first to see what each contributes.

Write four questions: two with source terms and two paraphrases. Predict passages, search manually, and record which rewrites help without changing intent.

Two comparisons

Renewal duration and pushing the return date back need C. Current quantity needs B; the 2025 limit needs E. One time condition changes the expected evidence.

Further reading: Microsoft: Vector search overview

← Previous · Learning plan · Fictional library practice sources · Next →