multi-agent-a2a-system

# 🧠 Multi-Agent AI System using Google's A2A Protocol This project demonstrates how to build a modular multi-agent AI assistant inspired by Google’s **Agent-to-Agent (A2A)** Protocol. Each agent handles a specific task—translation, memory retrieval, real-time web search, and final LLM-based response generation—communicating over structured HTTP. --- ## 📐 Architecture Overview - **Router Agent** (`router_server.py`): Orchestrates the entire flow, detects language, and coordinates other agents. - **Translator Agent** (`translator_server.py`): Converts non-English text to English. - **Memory Agent** (`memory_server.py`): Retrieves relevant knowledge using Graphiti + Neo4j. - **Search Agent** (`search_server.py`): Uses SerpAPI to fetch real-time web data. - **Final Agent** (`final_server.py`): Uses an LLM to generate the final, contextual response. - **Test Client** (`test_client.py`): Sends sample queries to the system. All agents follow the **A2A spec** by exposing a standard `/tasks/send` endpoint with structured JSON. --- ## 🚀 Setup Guide ### 1. Clone the Repository ```bash git clone https://github.com/manavpatel571/multi-agent-a2a-system.git cd multi-agent-a2a-system ``` --- ### 2. Install Dependencies Create and activate a virtual environment: ```bash python -m venv a2a-env source a2a-env/bin/activate # On Windows: a2a-env\Scripts\activate ``` Install required packages: ```bash pip install -r requirements.txt ``` --- ### 3. Configure Environment Variables Create a `.env` file in the root directory and add the following: ```env # OpenAI Configuration OPENAI_API_KEY=your_openai_api_key_here OPENAI_MODEL=gpt-4o OPENAI_EMBEDDING_MODEL=text-embedding-3-small # Google Cloud APIs GOOGLE_API_KEY=your_google_api_key_here CUSTOM_SEARCH_API_KEY=your_custom_search_api_key_here # SerpAPI SERPAPI_API_KEY=your_serpapi_key_here # A2A Agent URLs TRANSLATOR_URL=http://localhost:5001/tasks/send MEMORY_URL=http://localhost:5002/tasks/send SEARCH_URL=http://localhost:5003/tasks/send FINAL_URL=http://localhost:5004/tasks/send ``` > 💡 You can get a free SerpAPI key at: [https://serpapi.com/manage-api-key](https://serpapi.com/manage-api-key) --- ### 4. Start the Agents Open **5 terminals** (or use a process manager like `tmux`, `concurrently`, or Docker) and run: ```bash python translator_server.py # Port 5001 python memory_server.py # Port 5002 python search_server.py # Port 5003 python final_server.py # Port 5004 python router_server.py # Port 5006 ``` --- ### 5. Test the System Run the test client: ```bash python test_client.py ``` You should see the final AI response generated by chaining all agents. --- ## 📎 Notes - If SerpAPI fails (e.g., invalid key or rate limit), fallback logic returns a static placeholder message. - Neo4j and Graphiti setup is required for full Memory Agent functionality. - You can replace the LLM in `final_server.py` with GPT-4, Gemini Pro, or a HuggingFace model. --- ## 📚 Inspired By - [Google Cloud’s A2A Protocol Announcement](https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-agent2agent-protocol-a2a) - Concepts from LangChain, Neo4j, SerpAPI, and Flask-based microservices. --- ## 🛠️ Future Improvements - Add Docker support for one-click startup - Replace fallback search with DuckDuckGo or Bing - Add streaming and push notification support as per full A2A spec - Integrate authentication for agent discovery and invocation