Hello,
I have FAQs agent that reads PDFs stored in datastores. The agent was working fine, and I saved several examples, but it suddenly started to hallucinate and provide inaccurate answers to all the questions (even the ones saved as examples). I didn't make any changes, so I don't understand what could have happened.
Do you have any idea about the reasons?
Any help is appreciated
Hi. Is the agent using the tool for accessing the datastore?
Yes, the subagent invocations and datastore calls are working fine (I checked in the generated snippets).
It appears that the problem lies in the generation of the model's responses, which doesn't make any sense because it fails with the same questions that are already saved as examples.
What do you mean by fails? A noty shows up saying `Failed to generate response`? If that's the case maybe you added to many examples.
No, I don't get that notification.
For example, if I ask for the topic of example A, it responds with the topic of example B or with information that is not stored as an example but is in the datastore.
It's not matching the question with the content. However, I don't understand why, because I didn't make any changes.
@Myos wrote:Yes, the subagent invocations and datastore calls are working fine (I checked in the generated snippets).
Sorry, I think I'm missing something. Are the snippets returned by the tool relevant to your question? If they are, the last step is failing because the model is not answering based on this correct tool output?
User | Count |
---|---|
2 | |
2 | |
1 | |
1 | |
1 |