Folders can be learned with /learn multiple times

[3coins]

trafficstars

Summary

If a directory has been already indexed with a previous /learn command, and is submitted for learning again, this proceeds without any warning. However, when the /ask command is used subsequently, the response contains all the extra context because of the duplicate document in the vector store.

Note: So, far only observed this issue with the anthropic:claude-v1.2 model. With openai:gpt-3.5-turbo, the LLM seems to work ok through the duplicate content.

[JasonWeill]

Is this still an issue? Do any other models demonstrate it?

[JasonWeill]

@3coins Please verify

[3coins]

Could not replicate the issue with Anthropic claude-v1.2, but noticed that the /learn command is duplicating the data in the vector store.

[JasonWeill]

I still see the behavior @3coins mentioned above. Not sure whether this is a bug we need to fix; although the UI looks unusual, it is normal to learn a folder more than once, since its contents may change.