Hey guys, I’m working on a workflow with AI agent. He has to sort tickets coming from Linear, thanks to their title. But I am facing issues. The agent can’t stop hallucinating, inventing Ids, even with a strong prompt… Do you have any tips for that? Also, i give him in input a json, but sometimes he’s not able to reach easily the input, telling me no input is provided.
Model : Claude-3.7-sonnet
No Memory
OuputParser On
Have you considered changing the model? As far as I know, and based on my personal testing, Claude 3.7 Sonnet often responds with significantly more than what the user requested…
In some cases, it even adds content outside the scope of the original prompt…
I suggest trying other models and also lowering the model temperature, which might help improve response relevance…
Thanks for your answer.
Yes i also tried gpt-4o-mini, but too may hallucinations as well, evem more that claude-3.7-sonnet