---
title: GAIA Agent with Deep Research
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: "5.25.2"
app_file: app.py
pinned: false
hf_oauth: true
hf_oauth_expiration_minutes: 480
license: mit
---

# GAIA Agent with Deep Research

An intelligent agent built with LangGraph for the GAIA Benchmark, featuring multi-source research capabilities.

## 🌟 Features

- **Multi-Source Research**: Combines Wikipedia, Web Search, and Academic papers
- **RAG Enhancement**: Uses vector database for few-shot learning
- **Tool Orchestration**: Smart tool selection based on question type
- **Optimized System Prompt**: Guides the agent to choose appropriate tools

## 🚀 Quick Start

### 1. Install Dependencies

```bash
pip install -r requirements.txt
```

### 2. Configure API Keys

Create a `.env` file:

```bash
HUGGINGFACEHUB_API_TOKEN=your_hf_token
TAVILY_API_KEY=your_tavily_key
```

### 3. Setup Vector Database

```bash
python setup_chromadb.py
```

### 4. Run the Agent

```bash
python app.py
```

## 📁 Project Structure

```
hf_submission/
├── agent.py                 # Main agent implementation
├── deep_research_tool.py    # Multi-source research tool
├── app.py                   # Gradio UI
├── system_prompt.txt        # Optimized system prompt
├── setup_chromadb.py        # Vector DB initialization
├── requirements.txt         # Dependencies
└── metadata.jsonl           # Training data
```

## 🔧 Configuration

### Choosing LLM Provider

Edit `agent.py` line 147:

```python
graph = build_graph(provider="huggingface")  # or "groq", "google"
```

### Adjusting Deep Research

Edit `deep_research_tool.py`:

```python
# Change search sources
WikipediaLoader(query=query, load_max_docs=2)  # Adjust max_docs
TavilySearchResults(max_results=10)             # Adjust max_results
ArxivLoader(query=query, load_max_docs=5)      # Adjust max_docs

# Change content length
"content": doc.page_content[:2000]  # Adjust truncation
```

## 🎯 How It Works

### Agent Architecture

```
START → Retriever → Assistant → Tools → Assistant → END
           ↓           ↓           ↓
      Vector DB    LLM Decision  Execute
```

### Tool Selection Strategy

The agent automatically selects tools based on question type:

- **Complex topics** → `deep_research` (multi-source verification)
- **Simple facts** → `wiki_search` (quick lookup)
- **Recent events** → `web_search` (latest info)
- **Academic topics** → `arxiv_search` (papers)
- **Calculations** → math tools (multiply, add, etc.)

### Deep Research Process

1. **Multi-Source Search**: Queries Wikipedia, Web, and Arxiv in parallel
2. **Deduplication**: Removes redundant information using content hashing
3. **Formatting**: Creates structured report with source attribution

## 📊 Example Usage

```python
from agent import build_graph
from langchain_core.messages import HumanMessage

# Build agent
graph = build_graph(provider="huggingface")

# Ask question
question = "What is quantum computing and its applications?"
result = graph.invoke({"messages": [HumanMessage(content=question)]})

# Get answer
print(result["messages"][-1].content)
```

## 🛠️ Advanced Features

### System Prompt

The optimized system prompt (`system_prompt.txt`) guides the agent to:
- Choose appropriate tools for different question types
- Prioritize `deep_research` for complex topics
- Cross-validate important information
- Cite sources when possible

### Vector Database (RAG)

Uses ChromaDB to store past questions and answers:
- Provides few-shot examples to the agent
- Improves accuracy on similar questions
- Automatically populated from `metadata.jsonl`

## 🎓 Educational Resources

This project includes comprehensive documentation:

- Architecture deep dive
- Tool calling mechanism explained
- Deep research tutorial
- Integration guide

(See full project repository for detailed docs)

## 📝 License

MIT

## 🤝 Contributing

This is an educational project for the GAIA Benchmark. Feel free to:
- Report issues
- Suggest improvements
- Share your results

## 📮 Contact

For questions or feedback, please open an issue on the repository.

---

**Built with**: LangGraph, LangChain, HuggingFace, ChromaD