[Refactor] Replace LangChain with LangGraph: Node/Edge Graph and Tool Call Integration

[karthik18495]

Summary

Refactor the RAG4EIC codebase to use LangGraph instead of LangChain for orchestrating the LLM workflow. LangGraph enables more flexible, graph-based workflows using nodes, edges, and tool calls. This will improve modularity, maintainability, and enable advanced RAG behaviors (e.g., conditional tool use, branching, multi-hop reasoning).

Requirements

• Replace LangChain chains/runnables with LangGraph graphs (nodes and edges):
  • Each node represents an atomic operation (e.g., LLM prompt, retriever, tool call, response formatting)
  • Edges control data flow and can branch based on conditions (e.g., decide whether to call a retriever or respond directly)
• Integrate tool calls (e.g., retrieval, calculator, custom tools) within the graph
• Support for:
  • User message input node
  • Decide node (determine if retrieval is needed)
  • Retriever node(s) (arxiv, indico, or both)
  • LLM response node
  • Tool-calling node (for advanced queries)
  • Output formatting node
• Ensure all existing RAG features are mapped to graph nodes/edges (e.g., feedback, trace logging, source attribution)

Implementation Plan

1. Research LangGraph

   • Review LangGraph documentation and examples
   • Understand node/edge API and tool call patterns

2. Design the Graph Structure

   • Draw out the flow as a directed graph:
     • Start → User Input Node → Decide Node
       • If retrieval needed → Retriever Node(s) → LLM Node → Output Node
       • If not → LLM Node → Output Node
     • Add tool call node for specialized tools (e.g., calculator, DB query)
   • Allow branching for arxiv/indico/both retrievers based on user selection
   • Add nodes for logging, feedback, and error handling

3. Refactor Chains to Graph Nodes

   • Convert each "chain" or function in LangChain to a node in LangGraph
   • Wire up edges for data flow, including conditional branches (if/else)

4. Integrate Tool Calls

   • Implement tool-calling pattern per LangGraph's API (e.g., for retrieval, calculators, or custom actions)
   • Register tools and expose them as nodes in the graph

5. Testing & Migration

   • Write unit/integration tests for each node and the overall graph
   • Ensure all features work as before (or better)
   • Migrate streamlit UI to use LangGraph backend

6. Documentation

   • Document the new architecture and graph structure
   • Add diagrams showing node/edge layout
   • Provide migration/usage instructions for contributors

References

• LangGraph GitHub
• LangGraph Docs
• LangChain to LangGraph migration guide
• Existing chains in LangChainUtils/LLMChains.py and pages/2_RAG-ChatBot.py

Future Improvements

• Add more tool call support (e.g., web search, citation generator)
• Enable multi-agent or multi-hop reasoning via graph branches
• Visualize the workflow graphically in documentation or UI