Structured Output
JSON, XML, YAML
The Problem: AI often returns free-form text, but your code needs structured data like JSON or specific formats. How can we get consistent, parseable output?
The Solution: Fill Out the Form
Structured Output prompting specifies the exact format AI should use for its response. It's like giving someone a form to fill out instead of asking them to write a free-form letter. Essential for Prompt Chaining where each step needs parseable data, and few-shot examples help the model learn the exact format.
Think of it like filling out a structured form:
- 1. Define schema: Specify fields: name, email, department
- 2. Provide example: Show the exact JSON structure expected
- 3. Request data: "Extract contact info from this email"
- 4. Get structured result: Clean JSON ready to parse
Where Is This Used?
- API Responses: Generating JSON for web applications
- Data Extraction: Pulling structured info from unstructured text
- Form Generation: Creating structured records from descriptions
- Code Generation: Specific function signatures or class structures
Fun Fact: Modern APIs like OpenAI and Anthropic now support "function calling" which guarantees valid JSON output! This eliminates parsing errors and makes AI much more reliable for production use.
Try It Yourself!
Use the interactive example below to see how specifying output format leads to consistent, machine-readable responses.
Structured Output
Get data in a predictable format for programmatic processing
Popular output formats:
- • Hard to parse programmatically
- • Format may vary
- • Requires NLP to extract
Prompt Template
Analyze the following text and return the result ONLY in JSON format.
Text: "{TEXT}"
Use the following structure:
{
"field1": "type and description",
"field2": "type and description"
}
Important: return ONLY valid JSON without additional text.response = client.chat.completions.create(
model="gpt-4",
response_format={ "type": "json_object" },
messages=[
{"role": "system", "content": "Return JSON only"},
{"role": "user", "content": "Analyze this review..."}
]
)response = client.messages.create(
model="claude-3-opus",
tools=[{
"name": "extract_data",
"input_schema": {
"type": "object",
"properties": {
"sentiment": {"enum": ["positive", "negative"]},
"score": {"type": "number"}
}
}
}],
tool_choice={"type": "tool", "name": "extract_data"}
)- • Always specify "return ONLY JSON without additional text"
- • Use JSON Schema for client-side validation
- • Provide an example of expected structure in the prompt
- • Use response_format or tool_use for guaranteed JSON
- • Handle parsing errors — model may not follow instructions
| Method | Reliability | Flexibility | Complexity |
|---|---|---|---|
| Prompt Instructions | ⭐⭐ | ⭐⭐⭐ | Low |
| JSON Mode | ⭐⭐⭐ | ⭐⭐ | Low |
| Function Calling | ⭐⭐⭐ | ⭐⭐⭐ | Medium |
| Structured Outputs API | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | Medium |
Frequently asked questions
How do I get reliable JSON output from an LLM?
Use the API's built-in JSON mode (OpenAI's response_format, Anthropic's tool_use), provide a JSON Schema in the prompt, always say 'Return ONLY valid JSON', and implement client-side validation with a retry mechanism for malformed responses.
What is the difference between JSON Mode and Function Calling?
JSON Mode guarantees valid JSON output but doesn't enforce a specific schema. Function Calling (tool_use) lets you define exact schemas with required fields and types, giving you both valid JSON and guaranteed structure. Use Function Calling when you need specific fields.
Can LLMs output XML, YAML, or CSV reliably?
LLMs handle JSON most reliably since it dominates training data. XML and YAML work well with clear examples. CSV is less reliable for complex data. For best results, always specify the exact format, provide an example, and validate the output programmatically.
How do I handle parsing errors in structured output?
Implement a retry strategy: catch JSON parse errors, send the malformed output back to the LLM with 'Fix this JSON', and retry 1-2 times. Also use try-catch blocks, validate against your schema, and log failures for monitoring.
Try it yourself
Interactive demo of this technique
Extract contact information from text and return it as JSON
Contact information:
- Name: Alex Johnson
- Company: DataFlow Inc.
- Email: a.johnson@dataflow.com
- Phone: +1 (415) 555-0123
- Address: 42 Main Street
{
"name": "Alex Johnson",
"company": "DataFlow Inc.",
"email": "a.johnson@dataflow.com",
"phone": "+1 (415) 555-0123",
"address": "42 Main Street"
}
Without specifying a schema, the model returns a dashed list — readable for humans but not machine-parsable. A JSON schema + "only JSON" instruction produce machine-readable output that can be used directly in code.
Lesson Quiz
1.Why is structured output (JSON, XML, etc.) important when using LLMs in applications?
Create a free account to solve challenges
3 AI-verified challenges for this lesson
Related recipes
This lesson is part of a structured LLM course.
My Learning Path