I think it is supposed to answer from Vector_store, which is tied to Assistants.
The response itself is working fine.
However, it has been happening in the past few weeks that files that are not in Vector_store are being answered as annotations.
To explain the details, the file has not been deleted from the Vector_store. We continue to reference files that were originally there.
Check the file ID included in the annotation,
When I check the file acquisition API, etc., respectively, it returns “no file found”.
This behavior has been observed both through the Assistants API and in the playground.
If anyone has insights into the cause of this issue or suggestions for resolving it, please let me know.
Conversations that involve use of external tools look like this:
user: how big is an encypholous?
assistant: to tools->file_search, “encypholous size”
tools: [3] wadget_list.txt “thus when we see how large a sample can grow…”
assistant: "I don’t know what that is.
- [delete vector store file]
user: max samples size, though, you’d know?
assistant: samples can grow to 33 units (citation->[3] wadget_list.txt)
Deletion of a file only affects the future returns from search, not the persistent chat memory and internals hidden in a thread.
A vigilant developer would ensure that all files are current, but we can observe that deleting an image file will destroy a thread to which it was attached earlier, so OpenAI doesn’t do that reasonable diligence.
I’m having the same issue. I’m not sure who create the caching, the assistant or the vector store. If I remove the vector store from the assistant, it doesn’t reference the files anymore, if I put it back, same thing, I get reference of deleted files.
ChatGPT doesn’t seems to have this issue. When deleting the file in a custom GPT, the reference and knowledge is gone.