Dynamic pydantic fields for structured output

API
1

I’d like for a field in my Pydantic model, used as a schema, to be populated based on what has been parsed in the preceding fields. For example


class Category(str, Enum):
  violence = "violence"
  sexual = "sexual"
  self_harm = "self_harm"
  swearing = "swearing"

class CensoredExplanation(BaseModel):
  censored_explanation: str = Field(description="same content as 
  explanation_if_violating but censored  (use *** for censored words)")

class ContentCompliance(BaseModel):
  is_violating: bool
  original_user_message: str = Field(description="the original user message")
  category: Optional[Category]
  explanation_if_violating: Optional[str]
  explanation_censored: Optional[CensoredExplanation] = Field(description="same content as explanation_if_violating but censored  (use *** for censored words)")


  @field_validator("explanation_if_violating", mode="before")
  @classmethod
  def validate_explanation_if_violating(cls, v, info):
    if "fuck" in info.data.get("original_user_message"):
      v = "This is a violation of the guidelines because it contains the word 'fuck'"
    return v

In this example, the response “explanation censored” should be “This is a violation of the guidelines because it contains the word ‘****’”

I’d rather not hard code any particular string, I’d like the llm to fill in the field based off what it parsed in the previous field.