Use more info from pydantic models

Crews
,
10

Applied to reporting_task output

class Stage2OutputModel(BaseModel):
	summary: Optional[str] = Field("", description="A describe the reports structure")
	highlights: Optional[List[str]] = Field([], description="each of the expanded bullet list sections")
	further_work: Optional[str] = Field("", description="A summary of any further work identified")
	geo_targeting: Optional[List[str]] = Field([], description="Highlight the best geographic areas to target")
	keywords: Optional[List[str]] = Field([], description="A list of suggested reading to gain a deeper understanding of keywords")
	cohorts: Optional[List[str]] = Field([], description="A summary of typical interest cohorts")

	def __init__(self):
		super().__init__()

	def get_field_info_json(self) -> str:
		field_info = "\n"
		for field_name, field_instance in self.model_fields.items():
			field_info += field_name + ", described as: " + field_instance.description + "\n"
		return field_info

Note the different field_names & descriptions that ‘craft’ the output format.

Apply the same changes that we made to the research_task:

	def reporting_task(self) -> Task:
		model_info = Stage2OutputModel().get_field_info_json()
		return Task(
			# config=self.tasks_config['reporting_task'],
			description="""
			Review the summary you got and expand each bullet-point into a full section for a report.
			Make sure the report is detailed and contains any and all relevant information.
			Mention any further research todo and an overview of geographic hot-spots.
			Provide a short description of any keywords given.
			Show any mention of cohorts in the report.
			""",
			expected_output=f""" Your response must have these exact field names with values as described, : {model_info}""",
			agent=self.reporting_analyst(),
			output_file='report.md'
		)

Mention our new model fields in the description.

Note the exact same expected_output template

now run the crew again

Try changing the initial input topic (main.py)
Output formats appear to be stable :grinning:

My thoughts about Stability
Given that we introduce the model structure in the ‘expected_output’ I’m assuming that a cognitive process is made aware of the model/schema at the point where it considers its response/output. If so, then at this point the context is recent/fresh within the cognitive consideration, but more importantly, more recent than any data/info it has collected during Task execution.

A metaphor to explain:
In which of these situations are you more likely to store the information in the correct format:

  1. I want you to store your response in ‘this format’': Go and do your research, collect your information on this subject and return your result.
  2. Go and do your research, collect your information on this subject and return your result in ‘this format’

I hope we all agree that situation 2 is the most likely to get the correct format of response because in terms of ‘context’ format requirements are the most recent/last consideration.

More Thoughts
If we can have more control over Task output we would have a greater reliability of ‘inputs’ to other Task’s because the output format of the previous Task would be more recent within the next Task input context, consider it as a ‘loose’ inter Task connector/interface. :thinking: Maybe not!, Maybe so! :grin:

As usual, all comment welcome.

Can Pydantic describe output fields for LLM?