Closed sznariOsmosis closed 1 month ago
It's normal. The longer the query is, the lower the recall rate is.
It's normal. The longer the query is, the lower the recall rate is.
What is the purpose of this strategy, after testing, it uses the 4 questions of the history to retrieve chunks, which is easy to cause false detections and missed detections,Because I noticed that the code defaults to a new question composed of 3 questions from the current problem and the historical question, which are concatenated by strings
It's related to multi-turn conversations. We're gona refine it soon.
Is there an existing issue for the same bug?
Branch name
main
Commit ID
main
Other environment information
No response
Actual behavior
The debugging discovery is currently based on the latest 4 problems in the history to retrieve, and the debugging found that the consequence of this is that sometimes the content cannot be retrieved, but when I use the latest problem to retrieve, I can get the retrieved chunk
Expected behavior
What was the purpose of using the 4 questions to search in the first place, and will this be a hidden danger?
Steps to reproduce
Additional information
No response