Structured Output: Standardizing the AI

Most people use LLMs to write poems or summaries. But as an AI Engineer, you use LLMs to feed other software.

• If you want to update a database, you don't need a summary. You need a JSON object.
• User Query: "Extract name and age from this text."
• Bad Response: "Sure, the name is John and he is 30." (Hard to parse).
• Good Response: {"name": "John", "age": 30} (Easy to parse).

1. The Stability Problem

LLMs are creative by nature. Even if you tell it "Output JSON only," it might add conversational filler: "Here is your JSON: {`{ ... }`}". This will crash your Python code. Structured Output forces the model to follow a strict schema.

---

2. Pydantic: The Python Standard

LangChain uses Pydantic to define what the output should look like. Pydantic is a library for data validation.

from pydantic import BaseModel, Field

class Person(BaseModel):
    name: str = Field(description="The person's full name")
    age: int = Field(description="The person's age in years")

---

3. Visualizing the Schema Valve

graph LR
    LLM[Raw AI Output] --> V[Schema: Name(str), Age(int)]
    V -->|Pass| J[Clean JSON Object]
    V -->|Fail| E[Retry / Error]

---

4. The with_structured_output method

The modern way to get JSON from any LangChain model is using the .with_structured_output() wrapper. It handles the prompt engineering and selection logic for you.

structured_llm = model.with_structured_output(Person)
result = structured_llm.invoke("John Doe is 45 years old.")

print(result.name) # It returns a REAL Python object!
print(result.age)

---

5. Engineering Tip: "JSON Mode"

Most major providers (OpenAI, Anthropic) have a specific JSON Mode internal to the model. LangChain automatically enables this when you use with_structured_output.

---

Key Takeaways

• Structured Output is required for any system-to-system integration.
• Pydantic is the primary way to define AI data shapes.
• Machine-readable results prevent API crashes caused by AI "Chattiness."
• .with_structured_output() is the modern standard method in LangChain.