LangGraph Research Assistant

Multi-Agent AI Research System · LangGraph + FastAPI + React · Railway + Vercel

About the Project

A fully deployed multi-agent research assistant built with LangGraph. The user enters a research topic, configures analyst settings, reviews AI-generated analyst personas, approves them, and then watches the graph execute a complete research workflow.

The system conducts parallel, search-backed analyst interviews and synthesises the results into a structured markdown research report. The architecture demonstrates nested graphs, human-in-the-loop checkpointing, fan-out / fan-in parallel execution, and real-time streaming of subgraph progress to the frontend.

LangGraph

LangChain

FastAPI

React

Docker

Railway

Vercel

DeepSeek

Nested

Graph Architecture

Outer orchestration + reusable interview subgraphs

Human

In-the-Loop

interrupt_before checkpoint with resume from SQLite

Send()

Parallel Execution

Fan-out N analyst interviews with LangGraph Send API

Graph Execution Flow

outer

create_analysts

Generate analyst personas with structured Pydantic output

outer

human_feedback

Pause for approval — user reviews or regenerates analysts

checkpoint

outer

conduct_interview × N

Fan-out via Send() — one subgraph per analyst in parallel

inner

ask_question

Analyst generates a question for the expert

inner

search_web + search_exa

Parallel Tavily + Exa retrieval to build answer context

inner

answer_question

Expert answers using retrieved context

inner

route_messages

Loop or stop based on turn count / sign-off

inner

write_section

Analyst writes a section memo from the full transcript

outer

write_report / intro / conclusion

Parallel synthesis nodes assemble the final report

outer

finalize_report

Single markdown report delivered to frontend

Key Features

Nested Graph Architecture

Outer orchestration graph manages the full pipeline while reusable interview subgraphs run one analyst workflow each — clean separation of concerns at every level.

Human-in-the-Loop Approval

LangGraph's interrupt_before pauses execution at the analyst review step. The user can approve or give feedback to regenerate personas before any expensive work begins.

Parallel Analyst Interviews

LangGraph's Send() API fans out one interview subgraph per analyst simultaneously. Each runs independently with its own state, reducing total research time significantly.

Dual Retrieval: Tavily + Exa

Each analyst interview uses both Tavily for broad keyword-based web search and Exa for semantic highlight retrieval — giving richer, more accurate source context.

SQLite Checkpointing

LangGraph state is persisted with SqliteSaver at every step. Sessions survive server restarts and page refreshes, and the graph resumes exactly where it paused.

Streamed Graph Progress

Backend streams LangGraph updates with stream_mode='updates' including nested subgraph events. The frontend shows live node activity down to search_web and write_section.

Full Tech Stack

LangGraph

Multi-agent graph orchestration

LangChain

LLM integration & tooling

FastAPI

Python REST backend

React + Vite

Frontend SPA

Docker

Containerised backend

Railway

Backend deployment

Vercel

Frontend deployment

DeepSeek

LLM via LangChain

Tavily

Web search API

Exa

Semantic search API

SQLite

Checkpoint persistence

Pydantic

Structured output & validation

Try it Live

Enter a research topic and watch a multi-agent graph build a full report in real time

Open Live Demo View on GitHub