Conversational RAG with LangGraph Agents

📘 Description

This project extends a Retrieval-Augmented Generation (RAG) pipeline into a stateful, conversational assistant using LangChain, LangGraph, and OpenAI’s GPT-4o-mini. Designed for the MSDS 442 course at Northwestern University, it builds on previous modules by introducing agent reasoning, memory persistence, and multi-step retrieval workflows.

This reinforces the principle that AI systems are most powerful when embedded into structured, state-aware workflows — not standalone black boxes. LangGraph offers a clean architecture to reason over conversational history, perform multi-step tool calls, and allow LLMs to act as autonomous decision-makers.

⸻

🔧 Features

• Memory Persistence: Tracks multi-turn conversations using LangGraph’s MemorySaver checkpoints.
• Chat State as Messages: Uses LangChain’s message schema (user, AI, tool) for clean session tracking.
• Agent Reasoning (ReAct): Implements LangGraph’s ReAct-style executor to dynamically decide when and how to retrieve and respond.
• Multi-Step Retrieval: Empowers the agent to issue follow-up queries without requiring further user input.
• Visualization: Supports generation of Mermaid-compatible control flow graphs to inspect reasoning structure.

⸻

💡 Key Insight

Agents are ideal for dynamic and ambiguous queries. This project illustrates how LLMs can be embedded in intelligent systems that retrieve, reason, and respond across evolving conversations — making them more like collaborative assistants than static tools.

🔗 View the source code on GitHub