Forcing Structured JSON Output in LiteLLM + OpenRouter (FIXED)

[ahmed0x77]

Hey everyone! I ran into this exact issue and found three working solutions

The Problem: OpenRouter models DO support structured outputs, but LiteLLM's supports_response_schema() returns False for them, causing:

• CrewAI and other frameworks to skip sending response_format
• "Unknown parameter: response_format.response_schema" errors
• Broken structured output functionality

Root Cause: LiteLLM doesn't recognize openrouter as supporting structured outputs + the OpenRouter adapter strips/rewrites nested schemas.

Here's how I fixed it (pick what works for you):

---

Solution 1: Library Fix for OpenRouter Transformation

For the LiteLLM maintainers: The real issue is in the OpenRouter adapter that strips the response_format. Fix it here:

# Inside OpenrouterConfig.map_openai_params(...) method:

def map_openai_params(self, non_default_params, optional_params, model, drop_params):
    # ...existing code...
    
    # Add this block to preserve response_format:
    if "response_format" in non_default_params:
        rf = non_default_params["response_format"]
        if isinstance(rf, dict):
            # force LiteLLM to send your response_format exactly as given
            mapped_openai_params["response_format"] = rf
    
    # ...existing code...

AND add OpenRouter to the global support list:

PROVIDERS_GLOBALLY_SUPPORT_RESPONSE_SCHEMA = (
    "openai",
    "anthropic", 
    "cohere",
    "ai21",
    "mistral",
    "openrouter",  # ← Add this line!
)

You need both changes - the global list fix alone won't work because the OpenRouter adapter still strips the schema.

---

Solution 2: No-Code Workaround (For Users Right Now)

Use extra_body to inject the proper OpenRouter format:

import json, litellm

# Your schema
output_schema = {
    "type": "object", 
    "properties": {
        "title": {"type": "string"},
        "tags": {"type": "array", "items": {"type": "string"}},
        "description": {"type": "string"}
    },
    "required": ["title", "tags", "description"]
}

# This bypasses all the schema checking
response = litellm.completion(
    model="openrouter/mistralai/mistral-small-3.1-24b-instruct",
    messages=[{"role": "user", "content": "Analyze this..."}],
    extra_body={
        "response_format": {
            "type": "json_schema",
            "json_schema": {
                "name": "structured_response",
                "strict": True,
                "schema": output_schema
            }
        }
    }
)

# Parse the structured response
result = json.loads(response.choices[0].message.content)
print(result)

Works immediately with any OpenRouter model that supports structured outputs!

---

Solution 3: Manual Library Patch (Advanced Users)

If you want to patch your local installation right now:

# Find the OpenrouterConfig.map_openai_params method and add:

if "response_format" in non_default_params:
    rf = non_default_params["response_format"]
    if isinstance(rf, dict):
        mapped_openai_params["response_format"] = rf

---

For CrewAI Users Specifically:

Use Solution 2 with your existing setup:

# Your existing LLM setup
mistralLLM = BaseLLM(
    model="openrouter/mistralai/mistral-small-3.1-24b-instruct",
    base_url="https://openrouter.ai/api/v1", 
    api_key=OPENROUTER_API_KEY,
    temperature=0.0,
    # Add this to force structured output:
    extra_body={
        "response_format": {
            "type": "json_schema",
            "json_schema": {
                "name": "your_schema_name",
                "strict": True,
                "schema": your_pydantic_model.model_json_schema()
            }
        }
    }
)

---

💡 LiteLLM Team: This needs both fixes!

1. OpenRouter transformation fix - stop stripping response_format dicts
2. Global support list - add "openrouter" so supports_response_schema() returns True

The transformation fix is the critical one - without it, even adding to the global list won't work because the schema gets stripped before sending to OpenRouter.

---

Hope this helps everyone dealing with this! The extra_body workaround is your best bet until the library gets patched. 🔥