How to raise/show the errors from assistant?

[ShahriyarR]

Hello dear all,

Are by default the exceptions suppressed in the assistants?
Surprisingly It took me a while to understand why the assistant had not saved the data to the database.
It turned out there was an exception inside the method_tool but it was not raised:

@method_tool(args_schema=TestPatternSchema)
def save_test_pattern(self, test_pattern: str, task_uuid: str, test_implementation: str) -> None:
    ...
    task = Task.objects.get(uuid=task_uuid)
    pattern = Pattern.objects.get(uuid=task.pattern.uuid)

Only after enabling the Django logger within the assistant and logging the error it was shown:

'NoneType' object has no attribute 'uuid

After logging the error the assistant continued to run until it hit the recursion limit.
Before try/except and logging the error, it was just continuing.

 Task launch.tasks.epic_project_ai_flow[630c709e-1e89-4c7c-9f6f-f507ee0bf6c7] raised unexpected: GraphRecursionError('Recursion limit of 25 reached without hitting a stop condition. You can increase the limit by setting the `recursion_limit` config key.\nFor troubleshooting, visit: https://python.langchain.com/docs/troubleshooting/errors/GRAPH_RECURSION_LIMIT')

Not sure but maybe there is some retry mechanism already activated for try/except block and it fails exactly 12 times then hit the recursion error

Maybe just a naive question, but do we have any flag to reraise the exceptions or any kind of environment variables to enable us to show the exception trace?

Expected default behavior:

• Immediate termination of the chain in case of any exceptions - effectively fail fast.
• Showing the full stack trace after failure.
• Specific Assistant exceptions to catch and reraise or just a flag to disable the exception suppressing - this can be for example Recoverable and UnRecoverable exception sets - which we can also retry manually or just directly fail(just an ideation).

[fjsj]

Yes, there's an automatic retry mechanism in the ToolNode used internally in Django AI Assistant. See:
https://github.com/langchain-ai/langgraph/blob/784821705b5134e36c3e346854b6ae33f22cc100/libs/langgraph/langgraph/prebuilt/tool_node.py#L137-L181

Fail fast could be the default behavior, but it isn't right now.

The fastest mitigation here is to expose the graph building that's here into methods that you can override. I think the best solution is to extract each internal class and function there: AgentState, setup, history, retriever into methods, then also extract the graph building, so devs can override each graph node.

We can solve that, but it will take a few days / weeks. Feel free to open a PR if you wish too.
Right now, your best solution is to override as_graph in your AI Assistant class and change the ToolNode call. Does that make sense?

[ShahriyarR]

Thanks for the reply :)
I will check it at the weekend, and update the thread)

[ShahriyarR]

One more time, I would like to say thanks for this wonderful library. :) It really helped me get started easily with building agents, especially with Django. This was a missing part of the community.
I am planning to send proposals for PyCONs and give a talk about django-ai-assistant in 2025.

[fjsj]

That's awesome. Feel free to send me the proposal for review. My email is flavio at vinta.com.br