Structured Output And Parsers

Integrator asks the model for JSON in the prompt and parses the text with json.loads, which breaks when the model wraps output in prose or code fences.

Use model.with_structured_output(MySchema) (Pydantic/TypedDict/JSON schema) so the framework coerces output to the typed object via the provider's native mechanism. The invoke result is the typed object, not raw text to hand-parse.

Integrator uses PydanticOutputParser but never injects parser.get_format_instructions() into the prompt, so the model emits free-form text the parser rejects.

When using an output parser (vs with_structured_output), insert parser.get_format_instructions() into the prompt so the model knows the exact schema, then parse the result. The format instructions and the parser must reference the same schema.

A prompt | model chain returns an AIMessage object where the integrator expected a plain string, breaking string concatenation downstream.

Append StrOutputParser() to extract the text content from the AIMessage as a string when the downstream consumer wants text. Without it, the chain output is a message object; choose the tail parser to match the consumer's expected type.