I have an app that queries the text in a vector store.
I create an assistant and specify the model.
My prompt instructs the model to answer only from the attached files.
I create a thread for the assistant and attach a vector store (as per the api documentation, using file_search, vector_store_ids). The store is usually comprised of a handful of documents, none very big (10-50 pages of text)
I attach a couple of messages asking questions about the content of the documents
I execute the thread.
When using gpt-4o the response is exactly as expected, giving accurate answers from the documents. When specifying 4.1 the response is clearly from training data, and it makes no reference to the attached store — I saw something similar a year or so ago when I had failed to attach a vector store id.
However, besides the model specified, all the api calls are identical.
This issue seems similar to this post from October last year, but in this case it was 4o that couldn’t see the vector store. There were a lot of views and comments, but no fix, unfortunately. Assistants API vector store files not visible for gpt-4o models