It is normal that the final response generated by AI is different. It might be a summary over previous ongoing thoughts.

The final response is created by forcing AI to call a taskComplete tool, as otherwise AI bots may give up prematurely generating incomplete response.

@robin

Thanks for the feedback. Is there a way to prevent the deletion of the previous feedback? I may be able to determine if it is "thinking" that is being generated at first, but I doubt it to be honest. There is a lot of useful information in the response that now disapears.

When I specifically instruct the AI user to not call taskComplete everything stays put

Maybe I can compare response length to determine which response should be used. But I doubt this also creates issues for other situations...

Maybe just being able to prevent the call to "taskComplete" might already be enough?

Without this, many models will just give up too early with empty or partial response. This is the reason why taskComplete tool call is added in the first place.

I understand that. I was thinking along the lines of a checkbox in the AI settings that enables/disables the "taskComplete" method.

But it doesn't really matter. If I explicitly disallow the taskComplete call it works fine, so I have a workaround.

@robin Looking at the code in OneDev and imagining the final prompt the model receives it could actually think that the taskComplete() tool should be called with a summary of what has been generated before (= "provide your final answer using taskComplete tool"). I am also using Mistral and I regularly see the good content being thrown away and only a summary of that content is shown as final answer. In my experience smaller models need very explicit prompts otherwise they "guess" wrong.

Also in PRs when adding AI with mistral as reviewer an "empty response" message is added as comment which indicates the model has called taskComplete() but didn't provide data. Maybe the streaming result before did had that data but only the taskComplete() tool call got an empty string.

LLM does tend to summarize previous responses when calling taskComplete tool. Will remove this tool and use the response instead. When getting empty response, telling LLM to retry seems able to solve most of the issues.

State changed as build OD-7256 (15.0.0) is successful