Closed hanwesh closed 1 year ago
This is a regression in conversation quality due to the recent change from the (deprecated) davinci model to gpt35-turbo. We need to tweak the prompt and/or do some post-processing to improve the quality of the search queries generated by gpt35-turbo.
The "ask" tab still works reliably, since it doesn't rely on gpt35-turbo to generate search queries.
Another option for you is to use this other sample, which uses cognitive search built into OpenAI: https://github.com/microsoft/sample-app-aoai-chatGPT/blob/main/README_azd.md
This is a regression in conversation quality due to the recent change from the (deprecated) davinci model to gpt35-turbo. We need to tweak the prompt and/or do some post-processing to improve the quality of the search queries generated by gpt35-turbo.
The "ask" tab still works reliably, since it doesn't rely on gpt35-turbo to generate search queries.
Another option for you is to use this other sample, which uses cognitive search built into OpenAI: https://github.com/microsoft/sample-app-aoai-chatGPT/blob/main/README_azd.md
you are right @pamelafox I can see the ASK still working reliably. so what is next on this? will someone tweak the prompt from the team? Thanks
There was a recent change by @srbalakr that attempted to improve the query generation. I also think the recent change by @pablocastro may have improved it as well.
But according to recently filed #428, it is still occasionally generating queries with quotes. @akashtalole Do you have the latest version checked out?
@pamelafox I have noticed that issue #428 occurs intermittently, even though I checked out the latest version today.
Thanks for confirming. We likely need to do additional prompt engineering or possibly post-processing.
I tried with this prompt line added in query_prompt_template Do not return any text inside '' or "" in the search query terms.
Even in earlier code added post processing to completion.choices[0].message.content.replace("\"", "")
Need to test above fix.
after get the latest version as today, the chat seems back to normal now. Let me do more testing. @pamelafox @akashtalole
I did a recent analysis with our current prompt between davinci and gpt-35-turbo, and I'm seeing very similar search queries generated, so I think we can close this issue. When the instruct model becomes available in the future, we'll probably try that out and see if the quality is similar/better.
Minimal steps to reproduce
Any log messages given by the failure
Expected/desired behavior
OS and Version?
azd version?
Versions
Mention any other details that might be useful