Multi-Research Agent

A research-focused multi-agent system built with LangChain.

The main purpose of this project is to perform research by finding information from multiple sources, extracting relevant content, generating a report, and having a critic review the result.

Workflow

User
  ↓
Search Agent
  ↓
Reader Agent
  ↓
Writer Chain
  ↓
Critic Chain
  ↓
Research Report

Components

1. Search Agent

Performs the initial web research and finds relevant sources for the given topic.

Output:

search_results

2. Reader Agent

Works with the search results and identifies useful sources.

It uses a URL scraping tool to extract deeper content from selected webpages.

Output:

scraped_content

3. Writer Chain

Combines the search results and scraped content to generate a structured research report.

Output:

report

4. Critic Chain

Reviews the generated report and provides feedback instead of directly accepting the Writer's output.

Output:

feedback

State

A shared state is used to pass information between different stages of the workflow.

State
├── topic
├── search_results
├── scraped_content
├── report
└── feedback

This allows each stage to access information produced by previous stages.

Tools

The project currently uses tools for:

• Web searching
• URL scraping

These tools allow the agents to retrieve external information instead of relying entirely on the LLM's existing knowledge.

Project Structure

Multi_Research_agent/
│
├── src/
│   ├── agents/
│   │   └── agents.py
│   │
│   ├── pipeline/
│   │   └── pipeline.py
│   │
│   └── tools/
│       └── tools.py
│
├── main.py
├── .env
├── .gitignore
├── pyproject.toml
└── README.md

Tech Stack

• Python
• LangChain
• LLM
• Web Search
• Web Scraping
• Multi-Agent Architecture

For more complex agent workflows, frameworks such as LangGraph and CrewAI can be used to manage agents, tools, state, and execution flow.

Example

A research question such as:

Why did NVIDIA become dominant in the generative AI era
while Intel and AMD failed to capture the same opportunity?

can go through the complete pipeline:

Research Question
       ↓
   Web Search
       ↓
 Relevant Sources
       ↓
  Source Scraping
       ↓
 Research Content
       ↓
Report Generation
       ↓
  Critical Review
       ↓
Final Research Report

Current Status

This project is currently a local prototype.

It has not been deployed because a multi-agent workflow can generate multiple LLM/API calls during a single research task, which can significantly increase API usage and cost.

Future Improvements

• Persistent storage
• Conversational research
• Better source ranking
• Parallel research agents
• Improved report evaluation
• Web interface
• Deployment
• Cost and token optimization

Author

Amirtha Ganesh R