Would Reversing Few-Shot Prompt Placement in chatreadretrieveread.py Improve Accuracy?

[baku2san]

In the chatreadretrieveread.py file, the few-shot prompt is currently applied to the query creation side, but not to the response build_message. Would reversing this placement improve accuracy? I am considering submitting a Pull Request and would appreciate your feedback.

Current Few-Shot Prompt Placement: Link to current few-shot prompt

Proposed Placement: Link to proposed placement

the few shots setting
proposed placement

[pamelafox]

Hm, the few shots are currently only to provide examples of query rewriting. If we added few shots to the question-answering, they would be different answers. Is that what you mean?
If so, yes, we could consider that. I'd want to see the result of evaluations to see if its worth the additional tokens, however.

[baku2san]

What kind of evaluation are you looking for?
Does the current test include an evaluation? If so, I would like to create a pull request with the test results in the swapped state. In that case, I will not address the test.jsonfile but will make corrections if the direction is correct.