🔌 Ideal AI - Universal LLM Connector

One Connector to Rule Them All

A production-ready, Open Source Python LLM Connector providing a unified interface for Text, Vision, Audio, Image & Video across 15+ providers (Ollama, OpenAI, DeepSeek, Anthropic, Qwen, Xiaomi, xAI, Alibaba, Baidu, Minimax, Zhipu AI (Z.AI), Google, Infomaniak, etc.).

Features dynamic model injection (add new providers at runtime without code changes) and native support for Smolagents & LangChain workflows.

✨ Features

• 🔗 Universal LLM Connector - One unified interface for 15+ providers (OpenAI, Ollama, Anthropic, DeepSeek, Google, Xiaomi, Baidu, Minimax, xAI, Zhipu AI, Alibaba, etc.).
• 🎯 Multi-Modal Powerhouse - Text, Vision, Audio (STT), Image Gen, Video Gen (Wan 2.1), Speech (TTS).
• 💉 Dynamic Model Injection - Register new models or providers at runtime without changing source code.
• 🤖 Agent & Workflow Ready - Native wrapper for Smolagents and fully compatible with LangChain / LangGraph.
• 🎙️ Native Voice Chat - Ready-to-use pipeline for full audio-to-audio interaction.
• 🛡️ Production-Grade - Robust error handling, async polling (for Video/Audio), and binary management.
• 💼 100% Open Source - Apache 2.0 License, free for commercial use.
• 📦 PIP-Installable - pip install ideal-ai

📺 See it in action

One Connector to Rule Them All. Watch the full demo (2.50 min).

🚀 Quick Start

Installation

pip install ideal-ai

Basic Usage

from ideal_ai import IdealUniversalLLMConnector
import os

# Initialize with your API keys
connector = IdealUniversalLLMConnector(
    api_keys={
        "openai": os.getenv("OPENAI_API_KEY"),
        "google": os.getenv("GOOGLE_API_KEY"),
        "anthropic": os.getenv("ANTHROPIC_API_KEY"),
    }
)

# Text generation
response = connector.invoke(
    provider="openai",
    model_id="gpt-4o",
    messages=[{"role": "user", "content": "Explain quantum computing simply."}]
)
print(response["text"])

# Vision (multimodal)
with open("image.jpg", "rb") as f:
    image_bytes = f.read()

analysis = connector.invoke_image(
    provider="google",
    model_id="gemini-2.5-flash",
    image_input=image_bytes,
    prompt="What's in this image?"
)
print(analysis["text"])

# Image generation
result = connector.invoke_image_generation(
    provider="openai",
    model_id="dall-e-3",
    prompt="A futuristic robot in a cyberpunk city"
)
# result["images"] contains base64 or URLs

🎯 Pre-Configured Providers & Models – Fully Extensible (Inject Any Provider/Model at Runtime)

The following providers are pre-registered in config.json for immediate use. Note: You can easily inject any other model or provider (OpenAI-compatible, Ollama, etc.) at runtime without changing the package code.

Provider	Text	Vision	Audio	Speech	Image Gen	Video Gen
OpenAI	✅ gpt-4o, 3.5, 5	✅ gpt-4o, 5	-	✅ tts-1	✅ dall-e-3	-
Google	✅ gemini-2.5	✅ gemini-2.5	-	-	-	-
Anthropic	✅ claude-haiku-4.5	✅ claude-haiku-4.5	-	-	-	-
Ollama	✅ llama3.2, r1, qwen3	✅ gemma3, llava, qwen3-vl	-	-	✅ flux2 (4b/9b), z-image	-
Alibaba	✅ qwen3-max, plus, turbo, qwen3.6-35b	✅ qwen3.6-35b	-	-	✅ qwen-image-max	✅ wan2.1-2.5
Infomaniak	✅ apertus-70b, mixtral	-	✅ whisper	-	✅ flux-schnell	-
DeepSeek	✅ V3, R1, V4-Pro, V4-Flash	-	-	-	-	-
Moonshot	✅ kimi-k2.5	✅ kimi-vision	-	-	-	-
Zhipu AI	✅ glm-4.7	✅ glm-4.7	-	-	-	-
Baidu	✅ ernie-3.5, 4.0	-	-	-	-	-
Perplexity	✅ sonar	-	-	-	-	-
Hugging Face	✅ gpt-oss-120b	-	-	-	-	-
MiniMax	✅ M2	-	-	-	-	-
Mistral	✅ Small 4	-	-	-	-	-
xAI	✅ Grok 4	-	-	-	-	-
Xiaomi	✅ MiMo-V2.5, V2.5-Pro	✅ MiMo-V2.5	-	-	-	-

📚 Advanced Usage

Adding Custom Models at Runtime

The power of Ideal AI is its extensibility. Add any model without modifying source code:

# Define your custom model configuration
custom_model = {
    "myprovider:custom-model": {
        "api_key_name": "myprovider",
        "families": {
            "text": "openai_compatible"  # Reuse existing recipe
        },
        "url_template": "https://api.myprovider.com/v1/chat/completions"
    }
}

# Initialize connector with custom model
connector = IdealUniversalLLMConnector(
    api_keys={"myprovider": "your-api-key"},
    custom_models=custom_model
)

# Use it immediately
response = connector.invoke("myprovider", "custom-model", messages)

Dynamic Model Injection

# Add model after initialization
connector.register_model(
    "provider:new-model",
    {
        "families": {"text": "openai_compatible"},
        "url_template": "https://api.example.com/chat"
    }
)

Audio Transcription

# Transcribe audio with Infomaniak Whisper
transcription = connector.invoke_audio(
    provider="infomaniak",
    model_id="whisper",
    audio_file_path="recording.m4a",
    language="en"
)
print(transcription["text"])

Speech Synthesis (TTS)

# Generate speech from text
audio_result = connector.invoke_speech_generation(
    provider="openai",
    model_id="tts-1",
    text="Hello, this is a test.",
    voice="nova"
)

# Save audio file
with open("output.mp3", "wb") as f:
    f.write(audio_result["audio_bytes"])

Video Generation

# Generate video with Alibaba Wan (async polling handled automatically)
video_result = connector.invoke_video_generation(
    provider="alibaba",
    model_id="wan2.1-t2v-turbo",
    prompt="A robot walking in a futuristic city",
    size="1280*720"
)
print(f"Video URL: {video_result['videos'][0]}")

🤖 Smolagents Integration

Perfect for building AI agents:

from ideal_ai import IdealUniversalLLMConnector, IdealSmolagentsWrapper
from smolagents import CodeAgent

connector = IdealUniversalLLMConnector(api_keys={...})

# Wrap for smolagents
model = IdealSmolagentsWrapper(
    connector=connector,
    provider="openai",
    model_id="gpt-4o"
)

# Use with any smolagents agent
agent = CodeAgent(tools=[...], model=model)
agent.run("Build a web scraper for news articles")

🦜🔗 LangChain & LangGraph Ready

Ideal AI fits perfectly into LangGraph nodes or LangChain workflows. No complex wrappers needed—just call it directly inside your nodes.

from ideal_ai import IdealUniversalLLMConnector
from langgraph.graph import StateGraph

connector = IdealUniversalLLMConnector(api_keys={...})

# Use directly in a LangGraph node
def chatbot_node(state):
    response = connector.invoke(
        provider="deepseek",       # Switch provider instantly!
        model_id="deepseek-chat",
        messages=state["messages"]
    )
    return {"messages": [response["text"]]}

# Build your graph...
workflow = StateGraph(dict)
workflow.add_node("chatbot", chatbot_node)

---

🏗️ Clean Architecture & Enterprise Patterns

ideal-ai is built to be modular. For production applications, you can easily wrap it in a Service Layer to centralize your AI logic.

This approach gives you absolute control to inject custom parsers, switch providers dynamically (e.g., using Ollama for local development and OpenAI for production), and keep your business logic clean.

The Pattern: AIService Wrapper

# src/services/ai_service.py
from ideal_ai import IdealUniversalLLMConnector
import os


class AIService:
    """
    Centralized Service for AI interactions.
    Use this layer to manage environment-specific logic (Dev vs Prod).
    """
    def __init__(self):
        # Initialize the engine once
        self._engine = IdealUniversalLLMConnector(
            api_keys={
                "openai": os.getenv("OPENAI_API_KEY")
            }
        )

    def chat_with_user(self, user_message: str) -> str:
        """
        Your app's simplified contract.
        Centralizes the decision of which model/provider to use.
        """
        # Logic: Use free local model for Dev, powerful model for Prod
        is_prod = os.getenv("ENV") == "production"
        provider = "openai" if is_prod else "ollama"
        model = "gpt-4o" if is_prod else "llama3.2"

        response = self._engine.invoke(
            provider=provider,
            model_id=model,
            messages=[{"role": "user", "content": user_message}]
        )
        return response["text"]

Benefits of This Pattern

• ✅ Separation of Concerns - Business logic stays clean
• ✅ Environment-Aware - Dev uses local models, Prod uses powerful APIs
• ✅ Provider Abstraction - Swap providers without touching your code
• ✅ Testable - Mock the service layer easily
• ✅ Maintainable - All AI logic in one place

---

🔧 Configuration System

Ideal AI uses a two-level configuration system:

1. Families (Recipes) - Define how to interact with API types
2. Models (Cards) - Define which family each model uses for each modality

All default configurations are stored in config.json and can be extended without touching Python code.

Custom Parser Example

If a provider's response format is non-standard:

# Define custom parser
def my_parser(raw_response):
    return raw_response["data"]["content"]["text"]

# Inject it
connector = IdealUniversalLLMConnector(
    parsers={"provider:model": my_parser}
)

🐛 Debugging

Enable debug mode to inspect payloads and responses:

response = connector.invoke(
    provider="openai",
    model_id="gpt-4o",
    messages=[...],
    debug=True  # Shows raw API calls and responses
)

📦 Installation from Source

# Clone repository
git clone https://github.com/Devgoodcode/ideal-ai.git
cd ideal-ai

# Install in development mode
pip install -e .

# Or build and install
pip install build
python -m build
pip install dist/ideal_ai-*.whl

🧪 Running Examples

Check the examples/ folder for comprehensive demos:

# Open demo notebook
jupyter notebook examples/demo_ideal_universal_connector.ipynb

The demo notebook covers 13 comprehensive capabilities in a structured, progressive order:

Step	Feature	Description
0️⃣	Installation	Quick setup via pip install -U ideal-ai
1️⃣	Text Generation Loop	Unified iteration over 15+ providers
2️⃣	Vision/Multimodal	Image analysis (Gemini, GPT-4o, Claude, Kimi, GLM, Qwen-VL)
3️⃣	Image Generation	Create art (DALL-E 3, Flux, Z-image, Qwen-Image)
4️⃣	Audio Transcription	STT with Infomaniak Whisper with auto-polling
5️⃣	Speech Synthesis (TTS)	Natural text-to-speech with OpenAI
6️⃣	Video Generation	Async video creation with Alibaba Wan (auto-polling)
7️⃣	Runtime Injection	Register custom models/providers on the fly
8️⃣	Conversational Memory	Multi-turn chat history across providers
9️⃣	AI Agents	Autonomous agents with smolagents
🔟	Custom Parsers	Handle proprietary API response formats
1️⃣1️⃣	Debugging Mode	Inspect raw API payloads & responses
1️⃣2️⃣	Interactive Testing Interface	Bonus: All-in-one graphical dashboard for all modalities
1️⃣3️⃣	Summary	Next steps & Acknowledgments

🔑 Environment Variables

Create a .env file or set environment variables:

OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=AI...
ANTHROPIC_API_KEY=sk-ant-...
ALIBABA_API_KEY=sk-...
INFOMANIAK_AI_TOKEN=...
INFOMANIAK_PRODUCT_ID=...
MISTRAL_API_KEY=...
XAI_API_KEY=...
XIAOMI_API_KEY=...
OLLAMA_URL=http://localhost:11434

🚀 Built-in Models (Extensible to Any Provider)

These models are pre-registered in config.json for immediate use. Remember: You are not limited to this list! You can inject any new model or provider at runtime.

Text Generation

• OpenAI: gpt-4o, gpt-3.5-turbo, gpt-5
• Google: gemini-2.5-flash
• DeepSeek: V3 (deepseek-chat), R1 (deepseek-reasoner), V4-Pro, V4-Flash
• Infomaniak: apertus-70b, mixtral
• Anthropic: claude-haiku-4-5
• Alibaba: qwen-turbo, qwen-plus, qwen3-max, qwen3.6-35b
• Ollama: llama3.2, qwen3:30b, deepseek-r1:8b, mistral-small
• Moonshot: kimi-k2.5, kimi-k2-0905-preview
• Zhipu AI: glm-4.7
• Baidu: ernie-3.5, ernie-4.0
• Perplexity: sonar
• Hugging Face: gpt-oss-120b
• MiniMax: MiniMax-M2
• Mistral: mistral-small-2603
• xAI: grok-4-0709
• Xiaomi: mimo-v2.5, mimo-v2.5-pro

Vision/Multimodal

• OpenAI: gpt-4o, gpt-5
• Google: gemini-2.5-flash
• Anthropic: claude-haiku-4-5
• Ollama: llama3.2-vision, llava, qwen3-vl:30b, gemma3
• Moonshot: kimi-vision (moonshot-v1-8k-vision-preview)
• Zhipu: glm-4.7
• Alibaba: qwen3.6-35b
• Xiaomi: mimo-v2.5

Audio Transcription (STT)

• Infomaniak: whisper

Speech Synthesis (TTS)

• OpenAI: tts-1

Image Generation

• OpenAI: dall-e-3
• Infomaniak: flux-schnell
• Alibaba: qwen-image-max
• Ollama: flux2-klein:4b, flux2-klein:9b, z-image-turbo

Video Generation

• Alibaba: wan2.1-t2v-turbo, wan2.2-t2v-plus, wan2.5-t2v-preview

---

📖 Documentation

For detailed API documentation, see:

• GitHub Repository
• Connector API - Full method signatures with docstrings
• Configuration Schema - Available families and models
• Examples - Working code samples

🤝 Contributing

Contributions welcome!

To add a new provider:

1. Add family configuration to config.json (or pass as custom_families)
2. Add model configurations using that family
3. Test with the demo notebook

No Python code changes needed for most additions!

📝 License

Apache License 2.0 - See LICENSE file for details.

👤 Author & Support

Gilles Blanchet

• 🛠️ Created by: IA-Agence.ai - Enterprise AI Architecture & Custom Integration.
• 🌐 Agency: Idealcom.ch
• 🐙 GitHub: @Devgoodcode
• 💼 LinkedIn: Gilles Blanchet

🙏 Acknowledgments

This project is a labor of love, built on the shoulders of giants. Special thanks to:

• 🤗 Hugging Face: For the fantastic Agents Course. It inspired me to create this connector to easily apply their concepts using my own existing tools (like Ollama & Infomaniak) without the hassle of writing wrappers.
• My AI Co-pilots & Mentors:
  • Microsoft Copilot: For the architectural breakthroughs (Families & Invoke concepts) and our late-night debates.
  • Perplexity: For laying down the initial code foundation.
  • Google Gemini: For the massive refactoring, patience, and pedagogical support in improving the core logic.
  • Kilo Code (Kimi & Claude): For the security testing, English translation, and PyPI publishing preparation.
• The Model Providers: Ollama, Alibaba, Moonshot, MiniMax, OpenAI, Perplexity, Hugging Face, DeepSeek, Google, Zhipu AI, Baidu, Apertus, Anthropic, LangChain and Infomaniak for their incredible technologies and platforms.
• The Open Source Community: For the endless passion and knowledge sharing.

Built with ❤️ and passion, inspired by the open source AI community's need for a truly universal, maintainable LLM interface.

The adventure is just beginning...

---

One Connector to Rule Them All 🧙‍♂️