Build an End-To-End Modular RAG LangGraph AI Agent with Web Search

Static RAG pipelines struggle when user queries require information beyond the ingested documents, such as current events or real-time web verification. Pure web-search agents, on the other hand, lack grounding in proprietary internal documents.

This project implements a hybrid AI agent using LangGraph, FastAPI, and Streamlit that dynamically orchestrates document retrieval and live web search based on user intent and contextual relevance.

System Architecture and Routing Workflow

The core of the system is a stateful decision graph that manages multi-step execution paths:

                     ┌──────────────────┐
                     │    User Query    │
                     └────────┬─────────┘
                              │
                              ▼
                     ┌──────────────────┐
                     │  Router Decision │
                     └────────┬─────────┘
                              │
             ┌────────────────┴────────────────┐
             ▼                                 ▼
   [Local Knowledge Path]              [Live Web Search Path]
             │                                 │
     (Pinecone Vector DB)             (Tavily Search API)
             │                                 │
             └────────────────┬────────────────┘
                              │
                              ▼
                 ┌──────────────────────────┐
                 │ Context Synthesis & Gen  │
                 │    (Groq / LLaMA 3)      │
                 └────────────┬─────────────┘
                              │
                              ▼
                       [Final Answer]

1. Stateful Routing: LangGraph evaluates the incoming prompt and determines whether the answer resides in the local vector store or requires external web search.

2. Document Retrieval (RAG): For internal knowledge queries, the system searches indexed embeddings in Pinecone generated by lightweight HuggingFace MiniLM models.

3. Live Web Search: When real-time data is needed, or if the user explicitly enables web search, the agent routes execution to the Tavily Search API.

4. Fast Inference Synthesis: Retrieved chunks or web results are compiled into an augmented prompt and passed to Groq (LLaMA 3) for fast, grounded answer generation.

Technology Stack Summary

Key Takeaways

• Conditional Routing: By shifting from rigid linear chains to cyclic graphs with LangGraph, the agent can inspect context, evaluate retrieval quality, and fall back to web search when document context is insufficient.

• Traceable Decision Making: Exposing agent state transitions through FastAPI and Streamlit gives users full visibility into whether the model answered from internal PDFs or live web results.