Structured data has predictable fields and types. When inputs and outputs already follow a clear schema, ordinary programming is usually faster, cheaper, and more reliable than an LLM. An agent becomes useful when it must interpret varied, unstructured input—but its answer may still need to become structured data before another system can use it. section 1.4.1 section 2.1.3
A schema turns a natural-language request into a data contract. In Python, a Pydantic BaseModel declares field names and types and validates values against them. Pass the model as response_format to request JSON matching that structure. A required name and email can be paired with an optional phone number that defaults to None. section 2.2.5
Listing 2.7 Structured output with Pydantic
from pydantic import BaseModel
from litellm import completion
class ExtractedInfo(BaseModel):
name: str
email: str
phone: str | None = None
response = completion(
model="gpt-5-mini",
messages=[{
"role": "user",
"content": "My name is John Smith, my email is john@example.com, and my phone is 555-1234."
}],
response_format=ExtractedInfo
)
result = response.choices[0].message.content
The extraction example shows the complete path: declare a Pydantic model, request a matching response, then obtain the result from the response message. section 2.2.5
Structured data serves different points in an agent workflow:
- Tool calls encode a selected function and its arguments rather than free-form prose. section 3.2.2
- Final answers expose typed fields to downstream code, avoiding parsing that depends on the model's wording. section 4.7
- Evaluation responses separate whether a problem is solvable, the final answer, and any reason it cannot be solved. section 2.4.2
- Memory records preserve a problem, approach, result, and possible error analysis so past attempts can be retrieved and used. section 6.4.2
A tool-calling agent can produce a typed final answer without parsing its final prose. Give the agent an output_type, create a final-answer tool whose input is that Pydantic model, and require tool use when structured output is needed. Treat the answer as complete only after that tool succeeds; then return its validated result. An unsuccessful call must not be mistaken for a finished answer. section 4.7.1 section 4.7.2
Customer-feedback result
Define sentiment as positive, negative, or neutral; confidence as a float; and key_phrases as a list of strings. After the agent analyzes a review, its result.output is a Pydantic object whose fields the calling code can access directly. section 4.7.3
