SynGraph: Deep Research Agent

Overview

Course project (TNM114) that builds an agentic research stack: a router detects follow-ups vs new requests, a supervisor delegates to parallel researcher agents, and reports are synthesized with citations. Uses Gemini 3 Pro for reasoning, Gemini 2.5 Flash for routing/summarization, and Tavily + Exa for web/semantic search.

Key Features

Supervisor-worker LangGraph with typed state and parallel researchers
Clarification interrupts for ambiguous requests before planning
Dual search (Tavily + Exa) with compression and source tracking
Follow-up handling to merge prior context with new queries
Report generation pipeline producing cited research briefs

Challenges

Routing between new research and follow-up answers reliably
Keeping research loops transparent with cited sources
Balancing depth vs speed across Gemini Pro and Flash roles

Solution

Implements typed graph nodes for route/clarify/brief/supervisor/report, integrates Tavily and Exa connectors, and separates model roles so heavy reasoning stays on Gemini Pro while fast routing/summarization uses Flash.

Outcome

Demonstrates reliable deep-research UX with human-in-the-loop clarifications, parallel subtopic exploration, and cited reports suitable for coursework assessment.

Project Information

Technologies

TypeScript

Next.js

LangGraph

Google Gemini

Tavily

Exa

MDX

View Live Project View Source Code