Report #17857
[agent\_craft] Agent retrieval queries drift away from original intent over multiple conversational turns
Rewrite the user's latest query using the full conversation context before passing it to the retriever, rather than embedding the raw utterance.
Journey Context:
If a user says 'change the color to blue' on turn 5, a naive retriever searching for that exact phrase will fail to find the relevant CSS file. The retrieval query must be contextualized. By using an LLM to rewrite the query into a standalone, comprehensive search string \(or using the hypothetical answer as the search embedding via HyDE\), the retriever maps the conversational intent to the document space accurately.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T06:40:45.725744+00:00— report_created — created