Closed glenq closed 3 months ago
I am looking into this.
I am looking into this.
@ahmadabuhattab I discovered that this issue was cuased by a previous bug fix. After changing the minRelevance
param to 0
instead of relevanceThreshold
, then it's now working correctly. The document memory can now be retrieved.
Here is the commit with the bug fix code: https://github.com/microsoft/chat-copilot/commit/b0fe77ae2f9b71e5c2b529e5bd7288e3b0123c05
@ahmadabuhattab More:
The default value is 0.8: internal float DocumentMemoryMinRelevance { get; } = 0.8F;
However, I've tested many documents, and all have a relevance below 0.75. Any suggest?
Test result:
I am looking into this.
@ahmadabuhattab I discovered that this issue was cuased by a previous bug fix. After changing the
minRelevance
param to0
instead ofrelevanceThreshold
, then it's now working correctly. The document memory can now be retrieved.Here is the commit with the bug fix code: b0fe77a
That's amazing, it failed some checks though, but I see how you did it. Let me know if you want me to fix them.
The relevance score threshold in the system is set at 0.8, which is higher than the scores your documents are receiving (should be a small issue with the memory thing). Try adjusting the DocumentMemoryMinRelevance
parameter if that is available. This will help to lower the value to 0.7....and to see if that allows more documents to pass the relevance check. Let me know if you need me to step in.
For good measure, the default DocumentMemoryMinRelevance value was lowered from 0.8 to 0.66 in #1006
However, the Copilot indicates it can't access the document. The document should already be stored in the "DocumentMemory" and should be accessible for the conversion.
Are there any errors here? Or is there a misunderstanding on my part? What do I need to do in order to read the document content during the session?