How to Use Pydantic for Structured LLM Output

Introduction

When using large language models (LLMs) for tasks that require structured output, the variability of the generated text becomes a major obstacle; the same prompt can yield different answers, making downstream processing unreliable. To address this, Pydantic offers a solution that lets developers define custom object types for LLM responses.

Pydantic Overview

Pydantic is a Python library for data validation and settings management. By creating a schema (model) that inherits from BaseModel, developers can enforce field types, default values, and validation rules, ensuring that LLM output conforms to a predefined structure.

In the LangChain ecosystem, the PydanticOutputParser builds on Pydantic to provide JSON‑style parsing of LLM output.

Example: Structured Output with OpenAI Key

The following code demonstrates the full workflow using ChatOpenAI as the model.

from typing import List
from langchain.output_parsers import PydanticOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_core.pydantic_v1 import BaseModel, Field, validator
from langchain_openai import ChatOpenAI

Instantiate the model: model = ChatOpenAI(temperature=0) Define the expected output schema:

class Output(BaseModel):
    setup: str = Field(description="Give me the list of all the name of the characters")
    punchline: str = Field(description="Give me the list of places if it is available")
    @validator("setup")
    def question_ends_with_question_mark(cls, field):
        if field[-1] != "?":
            raise ValueError("Badly formed question!")
        return field

Set up the parser and prompt template:

parser = PydanticOutputParser(pydantic_object=Output)
prompt = PromptTemplate(
    template="Answer the user query.
{format_instructions}
{query}
",
    input_variables=["query"],
    partial_variables={"format_instructions": parser.get_format_instructions()},
)
chain = prompt | model | parser
answer = chain.invoke({"query": ""})
print(answer)

Example: Structured Output without OpenAI Key (using Ollama)

When an OpenAI key is not available, the same approach works with a local Ollama model.

from langchain_community.llms import Ollama
from langchain.output_parsers import PydanticOutputParser
from pydantic import BaseModel, Field
from langchain_core.prompts import PromptTemplate

Define the model name and create the Ollama instance (the model must be installed locally):

model_name = "model_name"
model = Ollama(model=model_name)

Define the output schema:

class Output(BaseModel):
    names: list = Field(description="Give me the list of all the name of the characters")
    places: list = Field(description="Give me the list of places if it is available")

Set up the parser and prompt:

parser = PydanticOutputParser(pydantic_object=Output)
prompt = PromptTemplate(
    template="Answer the user query.
{format_instructions}
{query}
",
    input_variables=["query"],
    partial_variables={"format_instructions": parser.get_format_instructions()},
)
chain = prompt | model | parser
query = "The Fable of the Fox and the Crow"
result = chain.invoke({"query": query})
print(result.names)
print(result.places)

Key Components Explained

PromptTemplate : Creates a prompt that includes the format instructions generated by the parser and the user query.

LLMChain : Represents the sequence prompt → model → parser . The prompt is filled, sent to the LLM, and the raw text is parsed into the Pydantic model.

Chain Invocation : The invoke method runs the whole pipeline with the provided query.

Conclusion

PydanticOutputParser, a core component of the LangChain toolkit, bridges raw LLM text and structured, JSON‑like data. By defining explicit schemas and integrating them into a LangChain pipeline, developers can reliably extract meaningful information from generative models, turning free‑form text into organized data ready for downstream processing.