v0.4.42: Complete Fix for Gemini Video Validation

What's Changed

• Complete fix for Gemini video validation - consistent field naming (#37)

Bug Fixes (Completes v0.4.41 fix)

• Fixed inconsistent camelCase/snake_case in file path video processing
• File path inputs now correctly use inline_data and mime_type (snake_case)
• Ensures consistency with base64 input handling from v0.4.41
• Fully resolves Issue #35: Video input fails with Gemini

Important Note

v0.4.41 only included partial fix (PR #36). This release includes the complete fix with PR #37.

Credits

• Fix completed by @devin in PR #37
• Original fix by @devin in PR #36

Full Changelog: v0.4.41...v0.4.42

v0.4.41: Fix Gemini Video Content Validation

What's Changed

• Fixed Gemini video content validation error (#36)

Bug Fixes

• Video content now properly wrapped in types.Part objects with inline_data
• Prevents "Extra inputs are not permitted" Pydantic validation errors
• Maintains full compatibility with base64 and file path video inputs
• Preserves existing MIME type detection logic

Credits

• Fix implemented by @devin in PR #36

Full Changelog: v0.4.40...v0.4.41

v0.4.40

Release 0.4.40

• README: Add Gemini 2.5 Flash Image docs (text-to-image, edit, fusion), aliases, return_images usage, availability notes
• Version bump to 0.4.40 in pyproject.toml, setup.py, src/init.py
• PyPI: https://pypi.org/project/ai-proxy-core/0.4.40/
• PR #34: #34

Link to Devin run: https://app.devin.ai/sessions/ac6ad5798ea7497ba0758163a0425b10
Requested by: Elijah Arbee (@ebowwa)

v0.4.3 - Universal System Instruction Abstraction

✨ Enhancement Release

This release provides universal system_instruction abstraction across all AI providers, building on the fix from v0.4.2.

What's New

• Universal system_instruction: Same parameter now works seamlessly across all providers
  • OpenAI/Ollama: Automatically converts to system message in messages array
  • Gemini: Passes as native system_instruction parameter
  • Anthropic: Converts to system parameter
• Zero code changes needed: Use the same system_instruction parameter everywhere
• Backward compatible: All existing code continues to work

Example Usage

from ai_proxy_core import CompletionClient

client = CompletionClient()

# Same system_instruction works for all providers\!
response = await client.create_completion(
    messages=[{"role": "user", "content": "Hello"}],
    model="gpt-4",  # or "gemini-1.5-flash", "claude-3", "llama2"
    system_instruction="You are a helpful pirate. Speak like a pirate."
)

Installation

pip install --upgrade ai-proxy-core==0.4.3

Links

• 📦 PyPI Package
• 🐛 Related Issue #32
• 💻 Commit: 2ef184e

Full Changelog: v0.4.2...v0.4.3

v0.4.2 - OpenAI Provider Fix

🐛 Bug Fix Release

This release fixes a critical issue with the OpenAI provider that was causing failures when using the unified CompletionClient.

What's Fixed

• Issue #32: OpenAI provider no longer fails with unexpected keyword argument 'system_instruction' error
• Added parameter filtering to exclude Gemini-specific parameters (system_instruction, safety_settings)
• Ensures seamless cross-provider compatibility

Installation

pip install --upgrade ai-proxy-core==0.4.2

Testing

The fix has been verified with comprehensive tests included in test_openai_fix_v0.4.2.py

Links

• 📦 PyPI Package
• 🐛 Issue #32
• 💻 Commit: fff44dd

Full Changelog: v0.4.1...v0.4.2

v0.4.1: Explicit Model Selection for Image Generation

🔧 Breaking Changes

This release corrects the image generation API to require explicit model selection. No automatic fallback or model selection.

Migration Required

# Old (v0.4.0)
provider = GPT4oImageProvider(api_key="...")
response = provider.generate(prompt="...")

# New (v0.4.1)  
provider = OpenAIImageProvider(api_key="...")
response = provider.generate(
    model=ImageModel.DALLE_3,  # REQUIRED
    prompt="..."
)

✨ What's New

• Explicit Model Selection: Every request must specify which model to use
• Three Models Available:
  • dall-e-2: Multiple images, editing, 256x256 to 1024x1024
  • dall-e-3: Styles, HD quality, up to 1792x1024, revised prompts
  • gpt-image-1: Token pricing, 4K resolution (4096x4096), better instruction following
• Model-Specific Validation: Each model validates its specific parameters
• Token Usage Tracking: GPT-Image-1 returns token usage information

📦 Installation

pip install ai-proxy-core==0.4.1

🚀 Quick Start

from ai_proxy_core import OpenAIImageProvider, ImageModel

provider = OpenAIImageProvider(api_key="your-key")

# Must specify model explicitly
response = provider.generate(
    model=ImageModel.DALLE_3,
    prompt="Modern app icon",
    size="1024x1024",
    quality="hd"
)

🔍 Available Models

Model	Max Size	Features	Special
dall-e-2	1024x1024	Edit, Multiple images (n≤10)	Lowest cost
dall-e-3	1792x1024	Styles, HD quality	Best for art
gpt-image-1	4096x4096	Token pricing, Better instructions	Best quality

🐛 Fixes

• Fixed GPT-Image-1 response_format parameter handling
• Corrected model-specific size validation
• Improved error messages for invalid models

---

Full Changelog: v0.4.0...v0.4.1

v0.4.0: GPT-4o Image Generation

🎨 Image Generation Support

This release adds comprehensive image generation capabilities to ai-proxy-core using OpenAI's DALL-E 3 API, with an abstract provider pattern for future GPT-4o native generation.

✨ Features

• Image Generation Provider: New GPT4oImageProvider for DALL-E 3 image generation
• Multiple Sizes: Support for SQUARE (1024x1024), LANDSCAPE (1792x1024), and PORTRAIT (1024x1792)
• Quality Options: STANDARD and HD quality settings
• Style Options: VIVID and NATURAL generation styles
• Image Editing: Edit existing images with natural language prompts and optional masks
• Azure Support: Full Azure OpenAI integration with AzureGPT4oImageProvider
• Localization: Generate images with localized text for app internationalization
• C2PA Metadata: Extract content authenticity metadata from generated images

📦 Installation

pip install ai-proxy-core==0.4.0

With OpenAI support:

pip install "ai-proxy-core[openai]==0.4.0"

🚀 Quick Start

from ai_proxy_core import GPT4oImageProvider, ImageSize, ImageQuality

provider = GPT4oImageProvider(api_key="your-openai-key")

response = provider.generate(
    prompt="A modern app icon with turquoise background",
    size=ImageSize.SQUARE,
    quality=ImageQuality.HD
)

with open("icon.png", "wb") as f:
    f.write(response["image"])

🌍 Localized Generation

# Generate app icons with localized text
locales = {"en": "CleanShots", "ja": "クリーンショット"}

for locale, text in locales.items():
    response = provider.generate(
        prompt=f"App icon with text '{text}'",
        size=ImageSize.SQUARE
    )

📝 Notes

• DALL-E 3 may struggle with accurate non-Latin text rendering
• The provider uses an abstract pattern for easy extension to other services
• Full integration with existing BaseCompletions architecture

🔗 Links

• Documentation
• PyPI Package
• Issue Tracker

---

Full Changelog: v0.3.9...v0.4.0

What's Changed

• feat: GPT-4o Image Generation Provider v0.4.0 by @ebowwa in #31

New Contributors

• @ebowwa made their first contribution in #31

Full Changelog: v0.3.9...v0.4.0

ai-proxy-core v0.3.9

ai-proxy-core v0.3.9

Patch release with audio support improvements and docs.

Highlights

• Gemini Live WebSocket: Allow 16-bit PCM audio passthrough; non-PCM inputs return a clear message (WebM→PCM conversion not yet implemented).
• POST /chat/completions (Google/Gemini): Accept audio inputs via data URL audio_url and OpenAI-style input_audio objects (base64 + format).
• README updated to document the above features.

Pull Requests

• #24 Enable PCM audio passthrough for Gemini Live WS; accurate error for non-PCM (fixes #23)
  #24
• #25 docs: update README for Gemini Live PCM and /chat/completions audio support (refs #23)
  #25

Install

• PyPI: https://pypi.org/project/ai-proxy-core/0.3.9/
• pip install -U ai-proxy-core

Notes

• Tag: v0.3.9
• Thanks @ebowwa