Agentic RAG Workflow

This project implements an Agentic RAG (Retrieval-Augmented Generation) workflow designed for reading internal technical documentation. Currently, the public WixQA benchmark is used for testing and optimization. It employs a ReAct agent, Milvus for vector storage, and NVIDIA NIM for LLM and Information Retrieval services.

The core configuration is located at: examples/RAG/simple_rag/configs/wix_config.yml

• Agent Architecture: ReAct Agent (_type: react_agent)
• LLM: qwen/qwen3-next-80b-a3b-instruct (via NVIDIA NIM)
• Embeddings: nvidia/nv-embedqa-e5-v5 (via NVIDIA NIM)
• Reranker: nvidia/llama-3.2-nv-rerankqa-1b-v2 (via NVIDIA NIM)
• Vector Database: Milvus
• Memory: Mem0 (saas_memory)

1. wix_retriever_tool_with_rerank: Retrieves information from the Wix Help Center knowledge base with re-ranking capability using nvidia/llama-3.2-nv-rerankqa-1b-v2.
2. vllm_retriever_tool: Retrieves information about vLLM.
3. web_search_tool: Performs internet searches using Tavily.
4. code_generation_tool: Generates Python code.
5. Memory Tools: add_memory and get_memory for managing long-term user preferences.

1. Environment Setup: Ensure you are in the project root and activate the virtual environment:

   cd NeMo-Agent-Toolkit
   source .venv/bin/activate

2. Install Dependencies:

   uv pip install -e '.[langchain]'
   uv sync --all-groups --all-extras

3. Docker Services: Start the Milvus vector database:

   docker compose -f examples/deploy/docker-compose.milvus.yml up -d

   Note: You may need to bootstrap data into Milvus if not already present (e.g., using scripts/bootstrap_milvus.sh).

4. API Keys: Ensure your environment variables are set (e.g., in a .env file):

   • NVIDIA_API_KEY
   • TAVILY_API_KEY (if using web search)

To run the agent with a specific query:

nat run --config_file examples/RAG/simple_rag/configs/wix_config.yml --input "How do I change my site's color palette in ADI?"

To start a FastAPI server for the workflow:

nat serve --config_file examples/RAG/simple_rag/configs/wix_config.yml --host 0.0.0.0 --port 8001 --workers 4

API Documentation will be available at: http://localhost:8001/docs

You can test the running server with curl:

curl --request POST \
  --url http://localhost:8001/generate \
  --header 'Content-Type: application/json' \
  --data '{
    "input_message": "Can I start accepting payments on my site while my Wix Payments account is still under verification?",
    "use_knowledge_base": true
}'

To run the frontend UI:

1. Navigate to the UI directory:

   cd external/nat-ui

2. Install dependencies:

   npm ci

3. Start the development server:

   PORT=8000 npm run dev

4. Access the UI in your browser at http://localhost:8000.

To evaluate the workflow using the defined metrics (Accuracy, Groundedness, Relevance, Trajectory Accuracy):

nat eval --config_file examples/RAG/simple_rag/configs/wix_config.yml

To run the prompt optimizer (using Genetic Algorithm):

nat optimize --config_file examples/RAG/simple_rag/configs/wix_config.yml

For a complete list of commands and testing flows, refer to: NeMo-Agent-Toolkit/my_scripts/test_workflow.sh