การแยกวิเคราะห์และตรวจสอบผลลัพธ์แบบมีโครงสร้าง

LLMs return prose by default, but your code needs structured data: JSON, a list, a typed object. Parsing free text with regex is fragile.

This lesson covers getting reliable structured output from models.

The first step is simply instructing the model to produce a specific format. But instruction alone is not enough; models drift, add prose, or wrap output in markdown.

LangChain output parsers do two jobs: they generate format instructions to inject into the prompt, and they parse the model's response back into a structured object.

Define the shape you want with a schema (e.g. a Pydantic model). The parser turns it into instructions and validates the result against the fields and types.

Add the parser's instructions into your prompt template so the model knows exactly what structure to emit.

After the model replies, the parser converts the text into your typed object, raising an error if it does not match the schema.

Models occasionally produce invalid JSON. A retry/fixing parser detects the failure and asks the model to correct its own output, turning a hard crash into a recoverable step.

Many modern models support a JSON mode or function/tool calling that constrains output to valid structured data at the API level. When available, this is far more reliable than prompt instructions alone.

A value can be the right type but still wrong: a negative age, an empty required field. Add validators so business rules are enforced, not just the data shape.

Agents chain steps together, feeding one output into the next. If a step emits malformed data, the whole chain breaks. Structured, validated output is what makes multi-step agents dependable.

Putting it together:

• Define a schema for the data you need
• Inject format instructions into the prompt
• Prefer native JSON/tool mode when available
• Parse and validate, with a retry parser as a safety net

You learned to get reliable structured data from LLMs.

• Output parsers generate instructions and parse responses
• Schemas validate shape and types
• Retry parsers recover from malformed output
• Native JSON/tool modes are most reliable when available