Langgraph & LlamaIndex Agent Demo

This project demonstrates the usage of Langgraph and LlamaIndex agents, showcasing how they can make intelligent decisions based on input queries and call the appropriate tools when necessary.

Project Structure

The project is organized into two main folders: langraph and llamaindex, each containing a Jupyter notebook and utility files for their respective agents.

├── README.md
├── langraph
│   ├── LanggraphAgentDemo.ipynb
│   └── utils
│       ├── agent.py
│       ├── prompts.py
│       ├── router.py
│       └── tools.py
├── llamaindex
│   ├── ReActAgentDemo.ipynb
│   └── utils
│       ├── prompts.py
│       ├── retrieval.py
│       └── search.py
└── requirements.txt

Dependencies & Tech Stack

• Langraph & LlamaIndex: Core libraries for creating and managing agents and workflows. Langraph supports state-based workflows, while LlamaIndex enables ReAct agent decision-making and tool integration.
• OpenAI GPT-4o: Utilized for natural language understanding and question-answering capabilities.
• Azure Cognitive Search: Stores and retrieves Tesla-related financial documents to provide agents with specific data for complex queries.
• Tavily: Adds an additional layer of web-based data retrieval for external searches, enhancing the agent’s ability to answer broader questions.
• Environment Variables: Managed via a .env file to securely store API keys and configuration details.

For a detailed list of Python dependencies, refer to the requirements.txt file.

Langgraph Agent Demo

In the LanggraphAgentDemo.ipynb, we demonstrate a classifier agent using Langgraph to handle the following:

• Classification: The agent first determines if a question is general or if it requires additional information from an internet search or a vector database.
• Tool-Calling: If additional information is required, the agent invokes a tool-calling node that can either search the web or retrieve data from a vector store.
• Answering Questions: If the question is general, the agent uses a chatbot to provide an answer.

Components

• Agents: Three agents are created—classifier_agent, chat_agent, and tool_using_agent.
• Tools: Two tools are available—search_web_tool for web search and retriever_tool for vector store retrieval.
• Workflow: The workflow is set up in a state graph where the agent determines the path based on the question type.
• Memory Persistence: Memory is used to persist the state between graph runs using MemorySaver.

LlamaIndex ReAct Agent Demo

In ReActAgentDemo.ipynb, we create a generalized example of a ReAct Agent using LlamaIndex with decision-making capability:

• Classification: The ReAct agent first prompted to determine whether it can answer a question directly or needs external information.
• Tool-Calling: If necessary, the agent calls other tools to either search the web or retrieve data from an Azure Cognitive Search.
• Answering Questions: If the agent can answer directly, it responds without calling external tools.

Components

• Agents: The ReAct agent is created using ReActAgent.from_tools, with memory buffering and callback management.
• Tools: The tools available are search_tool, and retriever_tool.
• Callback & Memory Management: Token counting and debugging are handled by CallbackManager, and conversation history is stored in ChatMemoryBuffer.