Welcome to Agentic AI: AI that Acts

You are familiar with AI assistants like Siri or Alexa that respond to your questions. Agentic AI is a step beyond this.

Agentic AI is designed to take actions, not just react.

Consider this difference:

• Typical AI (like Siri): You ask "What is the weather?" and it replies. It waits for your questions.
• Agentic AI: You might say "Plan a beach trip next weekend." Agentic AI would then perform actions for you. It could:
  • Research beaches.
  • Check flight and hotel availability.
  • Compare prices.
  • Book reservations (with your approval).

The core idea is that AI can set goals and act to achieve them independently.

The Agentic Chatbot I built

I built an Agentic AI chatbot to demonstrate basic Agentic AI concepts.

This repository contains the code for that chatbot.

CHECK OUT THE AGENTIC CHATBOT HERE.

How My Chatbot Demonstrates Agentic Behavior

My chatbot acts as a basic agent by:

• Deciding how to answer:
  • When you ask a question, the chatbot determines the best way to find the answer.
• Accessing Information: It can automatically:
  • Read provided documents.
  • Check Wikipedia.
  • Search the internet.
• Agentic Action:
  • The chatbot does not just stop at finding information, but it also performs actions in your behalf.
  • The chatbot creates issues in Github based on the interactions it had with you and the instructions you gave.

Examples of Agentic AI Concepts

You might have heard of Agentic AI in examples like:

• Claude's "Computer Use": This lets you ask Claude to perform tasks directly on a computer, like ordering food online. Youtube Video
• OpenAI's "Operator": This AI can browse the internet to complete tasks for you, such as booking flights. Youtube Video

These are examples of AI designed to take actions to assist you.

Under the hood

This section provides a technical overview of the Agentic AI Chatbot I have built.

Core Framework: Langchain

The chatbot is built using Langchain. Langchain simplifies the process of creating complex AI agent workflows by providing tools and abstractions for:

• Integrating Language Models: Connecting to various language models like OpenAI, Google, and Groq.
• Tool Management: Defining and using tools to extend the agent's capabilities.
• Memory Management: Handling conversation history and agent state.
• Agent Logic: Structuring the agent's decision-making process.

Langchain serves as the foundational layer for my chatbot, orchestrating the different components of the chatbot to work together effectively.

graph LR
    A[Understand Information] --> B[Make a Plan];
    B --> C[Take Action - Use Tools];
    C --> D[Learn & Improve];
    D --> A;

Loading

Language Models (LLMs)

Large Language Models (LLMs) models are the "brain" of the chatbot, responsible for:

• Understanding User Input: Processing and interpreting your questions.
• Generating Text: Crafting responses, performing reasoning, and formulating actions.

You can select from different LLMs in the sidebar, including:

• OpenAI GPT-4o
• Meta Llama-3.1
• Google Gemma-2

Tools: Extending Agent Capabilities

To go beyond simply answering questions, the chatbot utilizes tool calling.

Tools provide the agent with access to external information and the ability to perform actions. The chatbot is equipped with the following tools:

• Document Retriever: This tool allows the agent to access and search the documents you provide in the "input_files" directory.

  • Retrieval-Augmented Generation (RAG): The agent uses these documents to inform its answers, ensuring responses are grounded in provided information.
  • Vector Store (FAISS): For efficient searching of documents, the chatbot uses a FAISS vector store. Documents are converted into numerical vector representations using embeddings.
  • Embeddings (HuggingFace): HuggingFace embeddings are used to create vector representations of text, enabling semantic similarity search within the document store.

• General Knowledge Access: This tool enables the agent to search and retrieve information from Wikipedia.

  • Wikipedia Retriever: Provides access to a vast repository of general knowledge for answering a wide range of queries.

• Internet Search: This tool allows the agent to perform real-time searches on the internet using the Tavily Search API. - Tavily Tool: Enables the agent to access current information and answer questions requiring real-time data.

graph LR
    A[User Query] --> B(Chatbot Agent);
    B --> C{Query Type?};
    C -- Type 1 (e.g., Document Info) --> D1[Document Retriever Tool];
    C -- Type 2 (e.g., General Knowledge) --> D2[Wikipedia Tool];
    C -- Type 3 (e.g., Weather) --> D3[Internet Search Tool];
    C -- No Tool Needed --> E[Answer Directly];
    D1 --> F(Tool Result 1);
    D2 --> G(Tool Result 2);
    D3 --> H(Tool Result 3);
    F --> I(Chatbot Agent - Answer Generation);
    G --> I;
    H --> I;
    E --> I;
    I --> J[User Answer];

Loading

Agent Workflow: ReAct Logic

The chatbot operates using a ReAct (Reason, Act, Observe) framework. This framework defines the agent's decision-making process:

1. Reason: The agent analyzes your question and determines the best course of action to answer it. This may involve deciding which tool to use, or if direct generation is sufficient.
2. Act: Based on its reasoning, the agent takes an action. This could be:
   • Using a tool (Document Retriever, Wikipedia, Internet Search) to gather information.
   • Generating a direct text response if it has enough information.
3. Observe: The agent observes the outcome of its action.
   • If a tool was used, it processes the information retrieved by the tool.
   • If it generated a response, it prepares to present it to the user.

This process is guided by a Prompt Template, which provides instructions to the agent on how to use tools, prioritize information sources, and format its responses.

Memory and Conversation History

To maintain context and have coherent conversations, the chatbot utilizes memory.

• MemorySaver: The MemorySaver component is used to store the history of your conversation with the chatbot.
• Contextual Conversations: This memory allows the chatbot to remember previous turns in the conversation, enabling it to understand follow-up questions and maintain context throughout the interaction.

Each conversation is assigned a unique ID, ensuring that chat history is correctly tracked and managed for each user session.

Agentic Action: GitHub Issue Creation

A key demonstration of the chatbot's agentic nature is its ability to create GitHub issues.

• Beyond Information Retrieval: Instead of simply providing information, the chatbot can take actions based on its responses.
• Issue Reporting: In this demonstration, the chatbot automatically creates a GitHub issue with your question as the title and its answer as the body.
• Action-Oriented AI: This feature highlights the potential of Agentic AI to perform tasks and initiate actions, moving beyond traditional question-answering systems.

Conclusion

The Agentic AI Chatbot is built using a combination of Langchain, Large Language Models, and specialized tools. Its architecture is designed to:

• Access and Process Information: From provided documents, Wikipedia, and the internet.
• Reason and Act: Using the ReAct framework to make decisions and take actions.
• Maintain Context: Remembering conversation history for more engaging interactions.
• Demonstrate Agentic Behavior: By performing actions like creating GitHub issues.