Does `Crew` Pydantic output require OpenAI

General
1

The docs for Task say

Outputs a Pydantic model object, requiring an OpenAI client. Only one output format can be set.

Does that mean that I cannot use an opensource LLM and get a Pydantic output from a crew?

2

I am wondering the same. Although I set the Task with ‘output_pydantic=’ in a proper class I can only get a raw output. Any clue of how to use a different LLM and make it a Task to have a pydantic output?

3

Hello,

I have figured it out how to use different LLM and get pydantic output. Here is an example:

@task
 def task2(self) -> Task:
        return Task(
            description=".....",
            expected_output = "Expected output should be a pydantic model of Task2TaskOutput type with desc property set to list of dictionary"
            llm=OpenAI35kTurbo, # I am using Azure LLM
            callback=log_output, # you can skip this
            output_pydantic = Task2TaskOutput, # This is my pydantic Basemodel
        )

My pydantic output model looks like this:

class Task2TaskOutput(BaseModel):
    desc: list[dict[str, int]]

Three things that are absolutely important are:

  1. set llm in your Task. I know there is no attribite called llm in Task, but this works.

  2. Mention in your expected output that your are expecting a pydantic model. Play around with the words in expected output

  3. In agent, after you write the backstory, metion output format. For example

    backstory = You are a seasoned anomaly detection specialist adept at identifying unusual patterns and behaviors in datasets.

     Output format:

 ```"desc": list[dict[str, int]]```
1 Like
4

That is great, but it does not answer the question of whether you need an OpenAI model for this task, since that’s still an openAI model. I’ve recently had a very hard time to get structured output from both nvidia and ollama models. Both throw an error saying they don’t support structured output, while they somehow are able to do it with the openai and ollama models. Help would be greatly appreciated by anyone that knows more on how to get this working.