🤖 Intelligent Research Agent

A sophisticated AI-powered research agent that takes user prompts and generates comprehensive, structured research reports. Built with LangChain, LangGraph, and Hugging Face APIs.

🎯 Features

Core Capabilities

• Intelligent Prompt Analysis: Parses user prompts to identify main topics and subtopics using Chain-of-Thought reasoning
• Dual-Mode Research:
  • Detailed mode: Respects user-specified subtopics
  • Inference mode: Automatically identifies relevant subtopics when only a topic is provided
• Web Research Integration: Searches the web using DuckDuckGo for current, relevant data
• Advanced Prompting Techniques:
  • ✅ Chain-of-Thought prompting for logical reasoning
  • ✅ Few-Shot prompting for high-quality synthesis
  • ✅ Least-to-Most prompting for structured report generation
  • ✅ Role-based prompting for specialized perspectives
• LCEL Pipeline: Fully implemented LangChain Expression Language for composable chains
• Structured Report Generation: Produces professional, well-formatted reports
• Report Validation: Ensures output meets quality standards (2+ A5 pages, ~2000+ words)

Report Structure

Each generated report includes:

1. Executive Summary - High-level overview and key takeaways
2. Introduction - Background and research objectives
3. Detailed Analysis - In-depth coverage of each subtopic
4. Key Findings - Most important discoveries and insights
5. Implications & Recommendations - Practical applications and future directions
6. Conclusion - Summary and overall significance

🛠️ Architecture

System Components

ResearchAgent (Orchestrator)
    ↓
LangGraph Workflow (5 Nodes)
    ├─ analyze_prompt_node: Parse and structure user input
    ├─ research_node: Conduct web searches
    ├─ synthesize_research_node: Transform raw data into insights
    ├─ generate_report_node: Create structured report
    └─ validate_report_node: Quality assurance
    ↓
Advanced Prompting Templates
    ├─ Prompt Analysis Template (Chain-of-Thought + Role)
    ├─ Research Synthesis Template (Few-Shot)
    └─ Report Generation Template (Least-to-Most)
    ↓
Hugging Face LLM (Mistral-7B)
    ↓
WebSearcher (DuckDuckGo)

State Management

ResearchState = {
    "user_prompt": str,        # Original user input
    "topic": str,              # Identified main topic
    "subtopics": list[str],    # All subtopics to research
    "prompt_analysis": dict,   # Structured analysis output
    "research_data": dict,     # Web search results
    "structured_report": str,  # Final output report
    "validation_result": dict  # Quality metrics
}

📋 Prerequisites

• Python 3.8+
• Hugging Face API token (free tier available at https://huggingface.co/settings/tokens)
• Internet connection (for web searches)

🚀 Quick Start

1. Installation

# Clone or navigate to the project directory
cd "C:\Users\HP\Desktop\Om_programming\AgenticAI\Research Agent"

# Install dependencies
pip install -r requirements.txt

2. Configuration

Create a .env file in the project root:

# Copy the example
copy .env.example .env

# Edit .env and add your Hugging Face API token
# HUGGINGFACEHUB_API_TOKEN=your_token_here

Get your token at: https://huggingface.co/settings/tokens

3. Basic Usage

from agent import ResearchAgent

# Initialize agent
agent = ResearchAgent()

# Research with explicit subtopics
prompt = """
Research on Machine Learning with focus on:
- Deep Learning architectures
- Training methodologies
- Real-world applications
"""

result = agent.research(prompt)
print(result["structured_report"])

4. Run Your First Research

python -c "from agent import ResearchAgent; agent = ResearchAgent(); result = agent.research('Artificial Intelligence'); print(result['structured_report'])"

📝 Usage Examples

Example 1: Detailed Topic with Subtopics

agent = ResearchAgent()
result = agent.research("""
Research Quantum Computing with focus on:
- Quantum bits and superposition
- Quantum gates and circuits
- Current quantum computers
- Applications and challenges
""")

Example 2: General Topic (Auto-Inference)

agent = ResearchAgent()
result = agent.research("What are the impacts of blockchain technology?")
# Agent automatically identifies relevant subtopics

Example 3: Access Detailed Results

result = agent.research("Your research prompt")

# Access different components
print(result["topic"])                    # Main topic
print(result["subtopics"])                # List of subtopics
print(result["prompt_analysis"])          # Analysis details
print(result["research_data"])            # Web search results
print(result["structured_report"])        # Final report
print(result["validation_result"])        # Quality metrics

🧠 Advanced Prompting Techniques Explained

1. Chain-of-Thought (CoT)

Used in prompt analysis to break down reasoning:

• Step 1: Extract main topic
• Step 2: Identify explicit subtopics
• Step 3: Infer missing subtopics
• Step 4: Validate relevance

2. Few-Shot Prompting

Applied in research synthesis with examples of good synthesis:

• Shows how to transform raw data into coherent insights
• Demonstrates professional writing style
• Provides templates for structured output

3. Least-to-Most Prompting

Implemented in report generation:

• STEP 1: Organize content into sections
• STEP 2: Enhance with insights
• STEP 3: Create conclusion
• STEP 4: Apply professional formatting

4. Role-Based Prompting

System prompts assign specific expertise:

• "You are an expert research analyst"
• "You are an expert research synthesizer"
• "You are a professional research report writer"

📊 Report Output

Reports are generated in text format with:

• Length: 2,000-2,500+ words (exceeds 2 A5 pages)
• Sections: 6 comprehensive sections
• Formatting: Professional headers, proper spacing
• Quality: Validated for content coverage and length

🔧 Configuration Options

Edit the code to customize:

# In agent.py - WebSearcher class
max_results=5  # Number of search results per subtopic

# In HuggingFaceModelManager class
"temperature": 0.7      # Creativity level (0=deterministic, 1=random)
"top_p": 0.9           # Diversity of output
"max_new_tokens": 1000 # Maximum generation length

# Model selection
repo_id="mistralai/Mistral-7B-Instruct-v0.2"  # Can change to other HF models

🌐 Supported Models

The agent is tested with:

• Primary: mistralai/Mistral-7B-Instruct-v0.2
• Alternative: meta-llama/Llama-2-7b-hf
• Alternative: tiiuae/falcon-7b

To use a different model, modify repo_id in HuggingFaceModelManager.__init__().

📈 Workflow Execution Flow

User Input
    ↓
[Node 1: analyze_prompt] → Parse topic and subtopics
    ↓
[Node 2: research] → Search web for each subtopic
    ↓
[Node 3: synthesize] → Transform raw data to insights
    ↓
[Node 4: generate_report] → Create structured report
    ↓
[Node 5: validate] → Check quality metrics
    ↓
Output Report + Metadata

🚨 Troubleshooting

Issue: "Hugging Face API token not found"

Solution:

# Add to .env file
HUGGINGFACEHUB_API_TOKEN=hf_xxxxxxxxxx

Issue: "No search results found"

Solution:

• Increase max_results in WebSearcher
• Try different search terms
• Check internet connection

Issue: Report too short

Solution:

• Increase subtopics (more comprehensive research)
• Adjust max_new_tokens in model configuration
• Modify report template to require more detailed sections

Issue: Slow execution

Solution:

• Reduce max_results in WebSearcher
• Use a smaller/faster model
• Reduce max_new_tokens

📦 Dependencies

Package	Purpose
langchain	Core LLM framework
langchain-community	Community integrations
langchain-huggingface	HF model integration
langgraph	Graph-based workflows
duckduckgo-search	Web searching
huggingface-hub	HF API access
transformers	LLM tokenization
pydantic	Data validation
python-dotenv	Environment management
requests	HTTP requests

🤝 Integration with Other Systems

The agent can be easily integrated as a library in any Python project:

# In your Python script
from agent import ResearchAgent

agent = ResearchAgent()
result = agent.research("Your research topic")

# Use the results
print(result["structured_report"])
print(result["topic"])
print(result["subtopics"])

📄 License

Open source for educational purposes.

📁 File Structure

Research Agent/
├── agent.py              # Main research agent
├── config.py             # Configuration management  
├── utils.py              # Utility functions
├── requirements.txt      # Python dependencies
├── .env.example          # Environment template
└── README.md             # This file

🎓 Educational Value

This project demonstrates:

• LangChain's LCEL (Expression Language)
• LangGraph for state management and workflows
• Advanced prompting techniques for better LLM outputs
• Integration with Hugging Face models
• Web data aggregation and synthesis
• Structured report generation

🔬 Future Enhancements

Possible improvements:

• Multi-document synthesis
• Source citation and attribution
• PDF report generation
• Interactive Q&A on generated reports
• Fact-checking and verification
• Multiple language support
• Real-time streaming output
• Custom report templates

💡 Tips for Best Results

1. Be Specific: More detailed prompts → better subtopics
2. Use Industry Terms: Technical prompts get better results
3. Specify Focus Areas: Listing subtopics guides research
4. Monitor API Usage: HF free tier has rate limits
5. Review Output: Validate reports match your needs

---

Created: December 2025
Framework: LangChain + LangGraph
Model: Hugging Face (Mistral-7B)
Search: DuckDuckGo API