It is the models, not the endpoint.
You can save yourself the cost of sending the same thing again for no reason and just cache a hash of the input and serve a prior response if you haven’t changed the embedded files.
The vector store search is also powered by an embeddings model with the same symptom, and you will not be using the exact same input vector to algorithmically returned ranked semantic results with the file search tool, either.
Or, look at it another way, if you didn’t like the last response, there’s always another that might go a different path…