fix(gemini): filter unsupported schema formats for Gemini API

[colesmcintosh]

Title

fix(gemini): filter unsupported schema formats for Gemini API

Relevant issues

Fixes #11427

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• I have added a screenshot of my new test passing locally
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Changes

Problem

Gemini API calls were failing when tool schemas contained unsupported format values like format: "email". The error message indicates that Gemini only supports enum and date-time formats for STRING types, while Vertex AI supports additional formats like email, uri, etc.

Both gemini/ and vertex_ai/ providers were using the same schema transformation logic, which didn't account for these API differences.

Solution

This PR implements Gemini-specific format filtering in the GoogleAIStudioGeminiConfig class:

1. Overrides the _map_function method to apply Gemini-specific schema filtering after the standard transformation
2. Removes unsupported format values like email, uri, etc. while preserving supported ones like date-time and enum
3. Works recursively on nested objects, arrays, and anyOf constructs
4. Preserves all other schema properties (type, description, validation rules, etc.)

Modified Files

• litellm/llms/gemini/chat/transformation.py:
  • Added _filter_gemini_unsupported_formats() method for recursive schema filtering
  • Overrode _map_function() to apply Gemini-specific filtering
  • Only affects the gemini/ provider; vertex_ai/ behavior is unchanged

Added Files

• tests/test_litellm/llms/gemini/test_gemini_format_filtering.py:
  • Comprehensive test suite verifying format filtering behavior
  • Tests for nested objects, arrays, and preservation of other properties

Example Usage

# This now works for both providers:
tool = {
    "type": "function",
    "function": {
        "name": "git_commit", 
        "parameters": {
            "type": "object",
            "properties": {
                "author": {
                    "type": "object",
                    "properties": {
                        "email": {"type": "string", "format": "email"}  # This format gets filtered for Gemini
                    }
                }
            }
        }
    }
}

# Works with gemini/ (format removed)
litellm.completion(model="gemini/gemini-2.5-flash-preview-05-20", tools=[tool], ...)

# Works with vertex_ai/ (format preserved)  
litellm.completion(model="vertex_ai/gemini-2.5-flash-preview-05-20", tools=[tool], ...)

The fix ensures backward compatibility and follows the principle of making the libraries "just work" without requiring users to modify their schemas for different providers.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
litellm	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Jun 9, 2025 2:37pm