Structured Outputs

Get LLM responses as valid JSON using JSON mode or strict JSON Schema. Works with all supported providers.

Force LLM responses to follow a specific JSON format. Use JSON mode for flexible extraction, or strict JSON Schema for guaranteed schema conformance.

JSON Mode

Set response_format to {"type": "json_object"} to instruct the model to output valid JSON. Always include a system prompt describing the expected structure.

from openai import OpenAI
import json

client = OpenAI(
    base_url="https://api.aicredits.in/v1",
    api_key="sk-your-key-here",
)

response = client.chat.completions.create(
    model="openai/gpt-4o-mini",
    response_format={"type": "json_object"},
    messages=[
        {
            "role": "system",
            "content": """Extract invoice details and return as JSON:
{"vendor": string, "amount": number, "currency": string, "date": string}""",
        },
        {
            "role": "user",
            "content": "TechCorp invoice, ₹13,500, 15 Jan 2025",
        },
    ],
)

data = json.loads(response.choices[0].message.content)
print(data["vendor"])   # TechCorp
print(data["amount"])   # 13500

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.aicredits.in/v1",
  apiKey: "sk-your-key-here",
});

const response = await client.chat.completions.create({
  model: "openai/gpt-4o-mini",
  response_format: { type: "json_object" },
  messages: [
    {
      role: "system",
      content: `Extract invoice details as JSON:
{"vendor": string, "amount": number, "currency": string}`,
    },
    { role: "user", content: "TechCorp invoice, ₹13,500, 15 Jan 2025" },
  ],
});

const data = JSON.parse(response.choices[0].message.content!);
console.log(data.vendor);  // TechCorp
console.log(data.amount);  // 13500

Structured Outputs (JSON Schema)

For strict schema enforcement, use "type": "json_schema" with a full JSON Schema definition. The model only outputs JSON that matches the schema exactly (supported on OpenAI GPT-4o and later):

structured_outputs.py

from openai import OpenAI
import json

client = OpenAI(
    base_url="https://api.aicredits.in/v1",
    api_key="sk-your-key-here",
)

response = client.chat.completions.create(
    model="openai/gpt-4o-mini",
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "invoice",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {
                    "vendor": {"type": "string"},
                    "amount": {"type": "number"},
                    "currency": {"type": "string", "enum": ["INR", "USD", "EUR"]},
                    "date": {"type": "string", "description": "ISO 8601 date"},
                    "items": {
                        "type": "array",
                        "items": {
                            "type": "object",
                            "properties": {
                                "description": {"type": "string"},
                                "quantity": {"type": "integer"},
                                "unit_price": {"type": "number"},
                            },
                            "required": ["description", "quantity", "unit_price"],
                            "additionalProperties": False,
                        },
                    },
                },
                "required": ["vendor", "amount", "currency", "date", "items"],
                "additionalProperties": False,
            },
        },
    },
    messages=[
        {"role": "user", "content": "Parse: TechCorp invoice, 3x server licenses at ₹4500 each, 15 Jan 2025"},
    ],
)

invoice = json.loads(response.choices[0].message.content)
print(invoice)

Response Healing

When JSON mode is enabled and a provider returns malformed JSON, AICredits automatically repairs the response before returning it. This handles:

Truncated JSON (missing closing } or ])
Markdown code fences wrapping the JSON (```json ... ```)
Extra text before or after the JSON object
Trailing commas and minor syntax errors

# Provider returned (truncated):
'{"vendor": "TechCorp", "amount": 13500, "items": [{"description": "Server'

# After response healing:
'{"vendor": "TechCorp", "amount": 13500, "items": [{"description": "Server"}]}'

Response healing is transparent — you receive valid JSON even when the provider's raw output was slightly malformed. When healing is not possible, the raw response is returned and you must handle the parsing error in your code.

Prompting Tips

Show an example in the system prompt — include a sample JSON object so the model has a concrete target format
Use snake_case keys — models handle them consistently across languages
Enumerate allowed values — for fields with a fixed set, list them: "status must be one of: pending, approved, rejected"
Set temperature: 0 for extraction tasks — reduces hallucinated field values when the answer is in the input

Provider Support

Provider	JSON Mode	Strict JSON Schema
OpenAI GPT-4o / GPT-4o-mini	Yes	Yes
OpenAI o1, o3	Yes	Yes
Anthropic Claude	Yes (via prompt)	Via tool calling
Google Gemini	Yes	Partial
DeepSeek	Yes	No
Mistral	Yes	No

For maximum portability, use JSON mode with a clear prompt schema. Strict JSON Schema mode is most reliable on OpenAI models; for other providers, combine JSON mode with precise system prompt instructions.

JSON Mode

Structured Outputs (JSON Schema)

Response Healing

Prompting Tips

Provider Support

On this page