Vision Model Parameter Fix for GPT-5 and O-Series Models

Vision Model Parameter Fix for GPT-5 and O-Series Models

Version: 0.233.201
Fixed in: 0.233.201
Issue: GPT-5 and o-series models failed vision analysis tests with "Unsupported parameter: 'max_tokens'" error

Problem

When testing Multi-Modal Vision Analysis with GPT-5 models (e.g., gpt-5-nano) or o-series models (e.g., o1, o3), the test would fail with:

Vision test failed: Error code: 400 - {'error': {'message': 
"Unsupported parameter: 'max_tokens' is not supported with this model. 
Use 'max_completion_tokens' instead.", 'type': 'invalid_request_error', 
'param': 'max_tokens', 'code': 'unsupported_parameter'}}

Root Cause

Both the vision test endpoint (route_backend_settings.py) and the image analysis function (functions_documents.py) were using the max_tokens parameter unconditionally:

response = gpt_client.chat.completions.create(
    model=vision_model,
    messages=[...],
    max_tokens=50  # ❌ Not supported by o-series and gpt-5 models
)

However, o-series reasoning models (o1, o3, etc.) and gpt-5 models require the max_completion_tokens parameter instead of max_tokens.

Solution

Dynamic Parameter Selection

Implemented model-aware parameter selection in both vision test and vision analysis functions:

# Determine which token parameter to use based on model type
vision_model_lower = vision_model.lower()
api_params = {
    "model": vision_model,
    "messages": [...]
}

# Use max_completion_tokens for o-series and gpt-5 models, max_tokens for others
if ('o1' in vision_model_lower or 'o3' in vision_model_lower or 'gpt-5' in vision_model_lower):
    api_params["max_completion_tokens"] = 1000
else:
    api_params["max_tokens"] = 1000

response = gpt_client.chat.completions.create(**api_params)

Detection Logic

Uses max_completion_tokens:

• All o1 models: o1, o1-preview, o1-mini
• All o3 models: o3, o3-mini, o3-preview
• All gpt-5 models: gpt-5, gpt-5-turbo, gpt-5-nano

Uses max_tokens (standard):

• gpt-4o models: gpt-4o, gpt-4o-mini
• Legacy vision models: gpt-4-vision-preview, gpt-4-turbo-vision
• GPT-4.1 and GPT-4.5 series

Case-Insensitive: Detection works regardless of model name casing (GPT-5-NANO, gpt-5-nano, O1-PREVIEW, etc.)

Files Modified

1. route_backend_settings.py

Function: _test_multimodal_vision_connection()

Changes:

• Added model type detection
• Dynamic API parameter building
• Conditional use of max_completion_tokens vs max_tokens
• Removed static max_tokens=50 parameter

Line: ~299-370

2. functions_documents.py

Function: analyze_image_with_vision_model()

Changes:

• Added model type detection
• Dynamic API parameter building
• Conditional use of max_completion_tokens vs max_tokens
• Removed static max_tokens=1000 parameter

Line: ~2974-3075

3. config.py

Version Update: 0.233.200 → 0.233.201

Testing

Functional Test

Created functional_tests/test_vision_model_parameter_fix.py to validate:

1. Vision Test Parameter Handling

   • Dynamic parameter building
   • Model detection for o-series and gpt-5
   • Correct parameter selection
   • Old static parameter removed

2. Vision Analysis Parameter Handling

   • Dynamic parameter building
   • Model detection for o-series and gpt-5
   • Correct parameter selection
   • Old static parameter removed

3. Model Detection Coverage

   • 16 test cases covering all model families
   • Case-insensitive detection
   • Correct parameter selection for each model type

Running the Test

cd functional_tests
python test_vision_model_parameter_fix.py

Expected Output:

🚀 Testing Multi-Modal Vision Analysis Parameter Fix
=================================================================
🔍 Testing vision test parameter handling...
   ✅ Vision test uses dynamic API parameter building
   ✅ Model detection for o-series and gpt-5
   ✅ max_completion_tokens for o-series/gpt-5 models
   ✅ max_tokens for other models
   ✅ Old static parameter removed
✅ Vision test parameter handling is correct!

🔍 Testing vision analysis parameter handling...
   ✅ Vision analysis uses dynamic API parameter building
   ...
✅ All vision parameter fix tests passed!

Impact

Before Fix

• ❌ GPT-5 models: Vision test failed with parameter error
• ❌ o1/o3 models: Vision test failed with parameter error
• ✅ GPT-4o models: Vision test worked
• ✅ Legacy vision models: Vision test worked

After Fix

• ✅ GPT-5 models: Vision test passes with max_completion_tokens
• ✅ o1/o3 models: Vision test passes with max_completion_tokens
• ✅ GPT-4o models: Vision test still works with max_tokens
• ✅ Legacy vision models: Vision test still works with max_tokens

User Experience

• Users can now select and test GPT-5 models for vision analysis
• Users can now select and test o-series models for vision analysis
• No breaking changes for existing deployments
• Automatic parameter selection based on model type

Technical Details

API Parameter Differences

Standard Vision Models (GPT-4o, GPT-4 Vision):

{
    "model": "gpt-4o",
    "messages": [...],
    "max_tokens": 1000,  # ✅ Supported
    "temperature": 0.7   # ✅ Supported
}

Reasoning Models (o1, o3, GPT-5):

{
    "model": "o1-preview",
    "messages": [...],
    "max_completion_tokens": 1000,  # ✅ Required instead of max_tokens
    # temperature NOT supported for reasoning models
}

Why Different Parameters?

Reasoning models (o-series, GPT-5) use a different API contract:

• max_completion_tokens: Limits the completion length only
• No max_tokens: This parameter is not supported
• No temperature: Reasoning models don't support temperature adjustment

Standard vision models use the traditional parameters:

• max_tokens: Limits both prompt and completion tokens combined
• temperature: Controls randomness in responses

Related Features

• Multi-Modal Vision Analysis (v0.229.088)
• Vision Model Detection Expansion (v0.229.089)
• Document Intelligence OCR Integration
• Enhanced Citations with Vision Data

References

• Azure OpenAI API - Chat Completions
• GPT-4o Vision Documentation
• O-Series Reasoning Models

Troubleshooting

If Vision Test Still Fails

1. Check Model Name: Ensure model deployment name matches expected patterns
2. Check API Version: Use 2024-02-15-preview or later for vision support
3. Check Region Availability: Not all models are available in all regions
4. Check Deployment Status: Ensure model is successfully deployed in Azure

If Wrong Parameter Used

The detection logic checks for:

• 'o1' in model name (case-insensitive)
• 'o3' in model name (case-insensitive)
• 'gpt-5' in model name (case-insensitive)

If your model name doesn't match these patterns but needs max_completion_tokens, contact support or adjust the detection logic.

Version History

• v0.233.201: Fixed parameter selection for GPT-5 and o-series models
• v0.229.089: Expanded vision model detection to include GPT-5 and o-series
• v0.229.088: Initial Multi-Modal Vision Analysis feature