I’m trying to use the Assistants API at an enterprise level. It’s simple; I give each user an assistant with the ability to use File Search with their files. What I do is provide the user with a list of uploaded files, and upon selecting them, I tell the assistant:
Only use the document: document.md (File ID: file-file-bdiqD7qj7kOl3uiUbQPMF3tI)
This method works well when there are few files in the vector store, but as it grows (not much, to 15 or 20 files) the behavior becomes unstable and it responds with nonsense answers.
My question is:
On the OpenAI website it says this:
*Known Limitations
We have a few known limitations we’re working on adding support for in the coming months:
- Support for deterministic pre-search filtering using custom metadata."
This change would be great because it would allow me to preselect the documents I want to consult. Is there any news about this feature?
In the meantime, what can I do? I’m thinking of creating a vector store for each new user interaction with the selected files, but it doesn’t seem like a clean solution.
Thank you.