Back to Works

LangGraph Research Assistant

2026

A multi-agent research system built with LangGraph that automates deep-dive research on any topic. It generates a panel of AI analyst personas, conducts parallel expert interviews grounded in web and Wikipedia search, and synthesizes findings into a structured report — with a human-in-the-loop checkpoint to guide the process.

How It Works

The graph runs in two phases separated by a human approval step:

Phase 1 — Analyst Generation

  • Given a topic, the graph generates AI analyst personas, each focused on a distinct sub-theme
  • The graph pauses for human feedback — approve or request changes
  • If feedback is provided, analysts are regenerated incorporating the feedback until approved

Phase 2 — Parallel Interviews & Report

  • Each analyst conducts an independent interview with an AI expert
  • Each turn triggers parallel Tavily web + Wikipedia searches to ground answers in real sources
  • All interviews run simultaneously via map-reduce (LangGraph Send() API)
  • After interviews complete, report sections are written in parallel, then combined into a final report with introduction, body, and conclusion

Key Patterns

Repository
GitHub
Stack
LangGraph, LangChain, OpenAI (gpt-4o-mini), Tavily, Wikipedia, Pydantic, LangSmith
Platform
LangGraph Studio (dev server)
flowchart LR %% Node Definitions START([START]) END([END]) A[create_analysts] B{human_feedback} C[[conduct_interview
xN Parallel]] subgraph WritingTasks [Parallel Writing Phase] direction TB W1[write_report] W2[write_introduction] W3[write_conclusion] end D[finalize_report] %% Connections START --> A A --> B %% Feedback Loop B -- feedback --> A %% Approval Path B -- approve --> C %% Branching to Parallel Writing C --> W1 C --> W2 C --> W3 %% Re-converging W1 --> D W2 --> D W3 --> D D --> END