AetherWeaver: Unified LLM Routing & Orchestration Platform

简体中文

AetherWeaver is a powerful, extensible platform for building advanced LLM applications. It provides a unified interface for multiple model providers, intelligent routing, RAG capabilities, and OpenAI-compatible endpoints.

✨ Key Features

• Multi-Model Support: Integrates with OpenAI, Google Gemini, Deepseek, Claude, and more.
• Intelligent Routing: Dynamically routes requests to the best-suited model based on task complexity and type (e.g., code, vision, reasoning).
• Retrieval-Augmented Generation (RAG): Supports multiple embedding models (OpenAI, Cloudflare) and vector databases (Qdrant, Upstash, Pinecone).
• OpenAI-Compatible API: Drop-in replacement for OpenAI's API, allowing seamless integration with existing tools.
• Observability: Built-in integration with Langfuse for detailed tracing and monitoring.
• Extensible Agents: Build powerful agents with custom tools like web search.

🚀 Quick Start

1. Clone the repository:

   git clone https://github.com/inoribea/AetherWeaver.git
   cd AetherWeaver

2. Install dependencies:

   yarn install

3. Configure environment variables:

   cp .env.example .env.local

   Fill in your API keys in .env.local. At a minimum, you need one model provider API key (e.g., OPENAI_API_KEY).

4. Run the development server:

   yarn dev

5. Open your browser and navigate to http://localhost:3000.

🛠️ Configuration

The application is configured through environment variables. Key options include:

• ANALYSIS_MODE: Set to rule_based (default) for fast routing or llm_enhanced for more accurate, AI-powered routing.
• EMBEDDING_PROVIDER: Choose between OpenAI and Cloudflare for document embeddings.
• QDRANT_URL / UPSTASH_VECTOR_REST_URL: Configure your vector database connection.
• LANGFUSE_SECRET_KEY: Enable monitoring by providing your Langfuse secret key.

For a complete list of variables, see docs/vercel-guide.md.

OpenAI-compatible Providers (Generic)

This project supports any OpenAI-compatible provider using the pattern <PREFIX>_API_KEY + <PREFIX>_BASE_URL (e.g., OPENAI, NEKO, O3, OPENROUTER).

• Select default provider via OPENAI_COMPAT_PROVIDER=<prefix> (case-insensitive).
• If not set, the resolver tries: OPENAI → NEKO → O3 → OPENROUTER → the first discovered pair in env.
• Official OpenAI does not require OPENAI_BASE_URL.
• You can also override per-route model pools with real model IDs defined in models-config.json (e.g., BASIC_MODELS, STRUCTURED_OUTPUT_MODELS).

Example:

# Choose default provider
OPENAI_COMPAT_PROVIDER=

# O3 provider
O3_API_KEY=your_o3_key
O3_BASE_URL=

# Route overrides (real model IDs from models-config.json)
BASIC_MODELS=Qwen/Qwen3-235B-A22B-search
STRUCTURED_OUTPUT_MODELS=Qwen/Qwen3-235B-A22B-search

OpenAI-compatible v1 endpoint auth (optional):

ENABLE_API_AUTH=true
LANGCHAIN_API_KEYS=user_key_1,user_key_2
# Callers must send: Authorization: Bearer <user_key>

See .env.example and docs/vercel-guide.md for more details.

📚 Documentation

• Project Overview: High-level summary of features and configuration.
• API Usage: How to use the different chat APIs.
• RAG & Embeddings: Details on configuring RAG.
• Smart Routing: In-depth explanation of the routing logic.
• Deployment Guide: Guide for deploying on Vercel.

🤝 Contributing

Contributions are welcome! Please open an issue or submit a pull request.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.