Part 1 of 6 · 1 min read
Overview
In this module you will
- Name five query transformations (rewrite, multi-query, HyDE, decomposition, routing) and say which failure each one targets.
- Make a follow-up question such as "what about Starter?" stand on its own.
- Explain why a made-up answer (HyDE) can retrieve better than the question, and when it backfires.
- Measure each transformation by kind of question, including its cost in model calls and searches.
- Decide, with evidence, when a transformation is not worth adding.
An agent does not type tidy search queries. It carries a conversation, so the question it needs answered is often "how do I fix it?" with the subject living three turns back. Users say "money back" where the documents say "refund". A single question often hides two: "what does E-4290 mean and how do I reset my password?" Every one of these makes the question a poor search key, even when the index is perfect. Query transformations repair the key before searching. They are cheap to try and easy to over-use: each one is usually a model call, so each adds latency and cost, and some make things worse.
About the examples
The corpus is the invented 18-passage help centre from module 3, and the 16 questions were written by me (four each of follow-up, vocabulary-gap, compound and plain). In a real system a language model performs each transformation from a prompt. Here, small rule programs imitate them (a table of synonyms, a pattern for "what about X?", templates for hypothetical answers) so the mechanism is visible, testable and free to run. A real model will be more flexible and also more error-prone. With 16 questions, one question is six percentage points: read the results as direction, not as a benchmark. No real model or service was called.