🤖 Multi-Agent Research & Report System

A production-grade AI research platform where 5 specialized agents collaborate to research any topic, generate a structured professional report, and automatically fact-check the output — all in avg 27 seconds.

Live Demo → huggingface.co/spaces/Sohel2309/multi-agent-research

---

The Problem

Researching any topic properly takes hours — searching multiple sources, reading articles, extracting key facts, analyzing patterns, writing a structured summary, and then verifying the information is actually accurate.

Most people either skip steps (shallow research) or spend 4–6 hours doing it manually.

This system automates the entire pipeline in under 90 seconds using 5 specialized AI agents — each handling one stage of the research process, with the final agent fact-checking everything before you see it.

---

How It Works

User Query
    ↓
🔍 Research Agent              →  Searches live web, extracts key facts
    ↓              ↘
📊 Analyst Agent    🔎 Extra Search Agent    ← run in PARALLEL (asyncio.gather)
    ↓              ↙
✍️  Writer Agent               →  Writes structured professional report
    ↓
✅ QA Agent                    →  Fact-checks report against source data
    ↓
Final Report + Verdict (PASS / PASS WITH WARNINGS / FAIL)

After the Research Agent finishes, the Analyst Agent and Extra Search Agent run simultaneously via asyncio.gather — one analyzes the findings while the other gathers additional statistics. The Writer Agent waits for both, then produces a richer report using all gathered context.

Each agent is powered by Llama 3.3 70B via Groq and orchestrated using LangGraph state machines with conditional routing and shared state management.

---

Benchmarks

Metric	Result
Avg end-to-end pipeline time	27.1s across 4 benchmark queries
QA unsupported claims flagged	avg 3 per report across 3 diverse topics
Parallel execution	Analyst + Extra Search fire concurrently via asyncio.gather
Topics tested	Quantum Computing · Remote Work · EV in India · Social Media

---

Agent Architecture

Agent	Role	Input	Output
🔍 Research Agent	Live web search + extraction	User query	Key facts & statistics
📊 Analyst Agent	Pattern analysis (parallel)	Research data	Insights & implications
🔎 Extra Search Agent	Additional data gathering (parallel)	Research data	Extra statistics & examples
✍️ Writer Agent	Report generation	Research + Analysis + Extra context	Structured markdown report
✅ QA Agent	Fact verification	Report + All sources	Verdict + flagged claims

---

Features

• 5-agent LangGraph pipeline with async parallel execution, conditional routing, and shared state
• Live web search via Tavily API — real-time data, not stale training knowledge
• Async parallel execution — Analyst and Extra Search agents run simultaneously via asyncio.gather; benchmarked at 27.1s avg end-to-end
• Automated fact-checking — QA agent cross-verifies every report claim against source data; flags avg 3 unsupported claims per report with PASS / PASS WITH WARNINGS / FAIL verdict
• Session history — every report saved to SQLite, accessible and downloadable from sidebar at any time
• 4-tab result view — Final Report / Research Data / Analysis / QA Review
• Markdown download — export any report as a .md file instantly

---

Tech Stack

Layer	Technology
Agent Orchestration	LangGraph 1.2.6
Async Parallel Execution	Python asyncio + asyncio.gather
LLM	Llama 3.3 70B via Groq API
Web Search	Tavily Search API
LLM Framework	LangChain 1.3.11
Frontend	Streamlit 1.58.0
Session Storage	SQLite
Deployment	Hugging Face Spaces

---

Run Locally

1. Clone the repository

git clone https://github.com/Sohel2309/Multi-agent-research-system.git
cd Multi-agent-research-system

2. Create and activate a virtual environment

python -m venv venv

# Windows
venv\Scripts\activate

# macOS/Linux
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Configure API keys

Create a .env file in the project root:

GROQ_API_KEY=your_groq_api_key
TAVILY_API_KEY=your_tavily_api_key

Both APIs offer free tiers — no credit card needed.

5. Run the application

streamlit run app.py

Open http://localhost:8501 in your browser.

---

Project Structure

multi-agent-research-system/
├── state.py          # Shared state definition (AgentState TypedDict)
├── tools.py          # Tavily web search wrapper
├── agents.py         # All 5 agent functions with async support
├── graph.py          # LangGraph pipeline with asyncio parallel execution
├── database.py       # SQLite session storage (CRUD)
├── app.py            # Streamlit UI
└── requirements.txt  # Pinned dependencies

---

Author

Sohel Bhongade B.Tech, IIT Indore

GitHub · Live Demo