Duo Chat: The UI initially displays the correct answer, which is then unexpectedly replaced by a completely unrelated answer.

Bug Summary

Screen_Recording_2024-07-02_at_12.23.15

Please see Detailed Bug Cause/Fix directly

Details:

• The model appears to receive two requests despite only one question being submitted.
• This results in an unrelated answer overriding the correct one in the UI.
• The issue might lead to hallucinations in the model’s responses.
• Slack discussion (GitLab internal)

Observations:

• The Rails app sends the question once, but then sends another request to the AI gateway with a different prompt related to technical documentation.
• This results in an inconsistent and incorrect response.

Investigation Findings:

• The problem persists when using a self-hosted model but not with Claude, the GitLab AI vendor.
• The discrepancy lies in an additional aiCompletionResponse GQL API call, indicating extra responses from the AI gateway.
• The self-hosted model invokes search_with_retry, generating an extra request, while the GitLab AI vendor does not.

Guides:

• A video demonstrates the issue, showing two requests sent when using the UI but only one from the IDE.
• Debugging efforts are ongoing, focusing on the additional requests sent from the UI.
• Request in LangSmith

Conclusion: The issue likely stems from the self-hosted model's receiving two requests.