Lecture Notes on Building Advanced Agents with Langra

Introduction to Langra and Research Agents

• Focus on building advanced agents using Langra.
• Specifically using Lang graph compatible with V2 of Lang chain.
• Aimed at creating a Research Agent for detailed responses rather than quick conversation.

Characteristics of Research Agents

• Multi-step AI agents provide in-depth, well-researched answers.
• Users are generally more patient for detailed responses.
• Capable of referencing multiple sources, including Google searches and archive papers.
• Designed to facilitate conversational interaction with users.

Overview of the Graph Structure

• Graph Overview:
  • S: Start Node
  • Oracle: Decision maker (LLM with prompt), manages the nodes:
    • RAG Search Filter
    • RAG Search Fetch
    • Archive Web Search
    • Final Answer
• The Oracle determines the flow based on user queries.

Graph-based Agent Frameworks

• Traditional agent frameworks like React encourage iterative reasoning and action steps.
• React breaks down LLM generation into reasoning and action steps, feeding observations back into the LLM.
• Moving to a graph-based approach provides more control and transparency.

Lang Graph Overview

• Lang Graph: A framework to build agents as graphs, providing fine-grained control.
• Supports a higher level of customization and understanding of agent logic.

Components Needed for Building Agents

1. Installation of Prerequisites:
   • Libraries for visualization and data handling.
   • Example libraries include Pine Cone, OpenAI, and Semantic Router.
2. Components:
   • Archive Paper Fetch: Returns abstracts for academic papers based on IDs.
   • Web Search: Provides access to Google search for general queries.
   • RAG Search Components: Access AI archive papers and filter results.
   • Final Answer: Custom format for outputting structured research reports.

Setting Up the Knowledge Base

• Use Pine Cone and a prebuilt dataset of AI archive papers.
• Build a knowledge base using OpenAI's embedding model for paper indexing.

Building the Agent State

• Agent State: Tracks current execution state through the graph.
• Contains input from users, chat history, and intermediate results.
• Includes an operator to add to intermediate steps without replacing them.

Creating Custom Tools

• Tools Creation:
  • Archive Paper Fetch: Retrieves and formats paper abstracts.
  • Web Search: Uses an API to return structured search results.
  • RAG Search and RAG Search with Filter: Access knowledge base for targeted information.
  • Final Answer: Structures the report into a usable format.

Implementing the Oracle

• The Oracle integrates the LLM and prompt components to manage decision-making.
• Prompt Structure:
  • Includes system prompts, chat history, and user input.
  • Maintains a scratchpad for intermediate actions and reasoning.

Running the Agent

• The router identifies which tool to use based on the Oracle's decisions.
• The Run Tool function executes the tools defined in the graph.
• The graph is constructed by defining nodes and edges between components.

Testing the Research Agent

• Input different queries to evaluate the responses and flow.
• Review the structured reports generated by the agent, including introductions, research steps, conclusions, and sources.

Conclusion

• Langra allows for building complex agent structures with transparency and customization.
• The graph-based approach is favorable for developing tailored research agents.
• Other libraries like Haack and Llama Index are exploring similar functionalities.

Final Thoughts

• The lecture summarized the process of building a research agent using Langra and emphasized the advantages of a graph-based structure.