Oct 18: First card. The picture turns a follow-up into a question you could search with no earlier message. Sources are LlamaIndex's chat engine guide and Azure AI Search's query rewrite guide. Azure marks that feature preview.
1 What it is
A support thread rarely repeats the subject. The next message is "What about weekends?" LlamaIndex's chat engine guide describes condense_question as looking at the chat history and rewriting the user message into a query for the index. Azure AI Search's query rewrite guide describes a different pass: the service sends the search text to a generative model and retrieves with the original query plus the alternatives. Microsoft marks that feature preview, and says preview features are not recommended for production. The picture is the follow-up case. It does not call either API.
Illustration. One follow-up, then the same turn written so it can be searched alone. The hits are written for this picture.
Search textWhat about weekends?
MissWeekday support is open from 09:00 to 17:00. Saturday mail waits until Monday.
MissThe refund window is 14 days.
The picture searches the follow-up as typed. It does not name support hours, so both chunks miss.
2 Why it matters
Cost
The rewrite is another model call before search. Azure's guide says the service sends the query to a generative model. LlamaIndex's condense-question mode also calls a model to rewrite, then queries the index. This card does not add a price for that call.
Quality
A follow-up that says "weekends" does not name the product or the policy. A standalone question can. Azure's guide says its rewrite can correct typos and expand a query with synonyms. This card does not claim every rewrite is a better query.
Risk
Azure's guide says the rewritten queries might not keep every exact term from the original, which matters when the query is a product code. A rewrite can also search for a subject the user did not ask about.
3 How to apply · try today
Take the last two messages from a real support thread. Write the follow-up as a question someone could paste into search with no earlier message. If it still says "that" or "it," it is not ready.
Common mistake
Searching the follow-up as typed. "What about weekends?" does not say which product, or that the subject is support hours.
Self-check
The user asked when weekday support is open, then asked "What about weekends?" What should retrieval search?
Support-docs answer bot uses this concept in one build. The set on this tab is still the note below.
Coming soon. One follow-up from the support-docs answer bot, rewritten so the hours doc can match, with the original wording kept beside it.
Coming soon. How LlamaIndex condenses a chat turn, and how Azure's preview rewrite adds alternatives. Azure says not to send the debug flag in production.
Questions
What is query rewriting?
Query rewriting turns a follow-up question into a search query that stands on its own.
Is Azure's query rewrite the same as a follow-up rewrite?
No. LlamaIndex's condense-question mode rewrites a chat turn into a query for the index. Azure's preview feature sends the search text to a generative model and retrieves with the original query plus alternatives.
Can a rewrite drop a product code?
Azure's guide says the rewritten queries might not contain all of the exact terms the original query had.