Remove markdown code block markers from LLM's response

[ddragosd]

@roribio and I debugged today why it takes so long (60s+) for a task to complete when using output_pydantic .

It turns out that it was caused by gpt-4o insisting to respond with block code markers when returning the JSON.

We tracked down the fix to the Converter throwing JSONDecodeError Error and going down the path of handle_partial_json , which, for a simple JSON (120 lines, 7KB) it takes 60s+ to complete.

[bhancockio]

Hey @ddragosd !

Thank you for creating this PR!

Root issue:

• The models defined in output_pydantic and output_json are not passed into the LLM calls. As a result, the outputs of a task do not match the specified output format which leads us to doing another LLM call to format the original LLM output.

Fix:

• When output_pydantic and output_json are present, append a stringified version of the schema to the prompt.

Improvement stats:

• When running a simple crew with 1 task and a output_pydantic , our old approach would take around 30 seconds to complete the crew kickoff
• Half of this time was dedicated to doing a followup LLM call to convert the output of the task into the specified pydantic model.
• Now, the outputs of tasks automatically generate the correct format so it only takes 15 seconds which is a 50% speed increase.
• We also do one less LLM call which saves users money as well

Closed by #1651

[ddragosd]

@bhancockio, I'm glad you caught a deeper issue with crewAI making two LLM calls instead of one, and improved the response time overall. This is awesome. I'm testing it right now. Thanks so much for the quick fix !!

Meanwhile, what do you think about support for [structured outputs] (https://openai.com/index/introducing-structured-outputs-in-the-api/)? I understand not all models have it today. But would you have any recommendations for how we could leverage this right now? I was gonna look for a way to support this, even work on a PR. Thanks !