Grounding Agent with Gemini AI

A powerful AI agent built with Google's Gemini API featuring:

• Two-stage pipeline: Grounding → Refinement
• Google Search grounding for factual accuracy
• 6+ custom tools for various operations
• Automatic function calling with seamless integration

Features

Two-Stage Architecture

1. Stage 1 - Grounding: Uses Gemini with Google Search and custom tools to gather factual, tool-augmented information
2. Stage 2 - Refinement: Polishes the response for clarity and presentation

Available Tools

1. calculator - Evaluate mathematical expressions

   • Supports: basic math, trigonometry, logarithms, square roots, etc.
   • Example: "Calculate sqrt(256) + 15 * 3"

2. get_current_datetime - Get current date and time

   • Returns: date, time, day of week, timezone info
   • Example: "What's today's date?"

3. get_weather - Weather information for any location

   • Supports: celsius/fahrenheit, humidity, wind speed
   • Example: "What's the weather in Tokyo?"

4. file_operations - File system operations

   • Operations: read, write, list, exists
   • Example: "List files in the current directory"

5. web_scraper - Extract content from web pages

   • Extract: text, title, metadata
   • Example: "Get the title of https://example.com"

6. text_analyzer - Analyze text content

   • Analysis: word count, sentiment, summary
   • Example: "Analyze sentiment: 'I love this product!'"

7. Google Search - Built-in web grounding

   • Automatically searches and grounds responses in real-time data
   • Example: "What was the score of the latest Olympic event?"

Installation

# Clone or download the repository
cd Web_agent

# Install dependencies
pip install -r requirements.txt

Usage

Interactive Mode

python main.py

This launches an interactive session where you can:

• Ask questions with natural language
• Use multiple tools in a single query
• Toggle Google Search grounding on/off
• View available commands with help

Programmatic Usage

from grounding_agent import GroundingAgent

# Initialize agent
agent = GroundingAgent(api_key="YOUR_API_KEY")

# Process a query
result = agent.process_query(
    "What's the weather in London and calculate 25% of 500?",
    use_search_grounding=True
)

# Display result
agent.display_result(result)

Example Queries

Multi-Tool Queries

"What's the weather in San Francisco and calculate 15 * 23?"
"What's today's date and what's the square root of 144?"

Web Search Grounding

"What are the latest developments in quantum computing?"
"What was the score of the latest NBA finals game?"

Text Analysis

"Analyze this text: 'I absolutely love this product! It's amazing!'"
"Count words in this text: 'The quick brown fox jumps over the lazy dog'"

Complex Calculations

"Calculate: (25 * 4) + sqrt(169) - 10"
"What's the sine of 45 degrees times pi?"

Configuration

API Key

The API key is currently hardcoded in main.py. For production:

API_KEY = os.getenv("GEMINI_API_KEY")  # Use environment variable

Model Selection

Change the model in grounding_agent.py:

agent = GroundingAgent(api_key=API_KEY, model="gemini-2.0-flash-exp")

Available models:

• gemini-2.0-flash-exp (default - fastest)
• gemini-2.5-flash
• gemini-pro

Project Structure

Web_agent/
├── main.py                 # Interactive CLI application
├── grounding_agent.py      # Core agent with two-stage pipeline
├── tools.py                # Custom tool definitions
├── requirements.txt        # Python dependencies
└── README.md              # This file

How It Works

1. Grounding Stage

def grounding_stage(self, query: str, use_search_grounding: bool = True):
    # Combines custom tools + Google Search
    # Gemini automatically calls relevant tools
    # Returns grounded response with sources

2. Refinement Stage

def refinement_stage(self, grounding_result: Dict[str, Any]):
    # Takes grounded response
    # Refines for clarity and structure
    # Returns polished final answer

3. Tool Integration

Tools are Python functions with docstrings. Gemini automatically:

• Parses function signatures and docstrings
• Determines which tools to call
• Executes tools and integrates results

Interactive Commands

When running main.py:

• help - Show available commands
• tools - List all tools with descriptions
• examples - Show example queries
• search on/off - Toggle Google Search grounding
• clear - Clear the screen
• quit or exit - Exit application

Response Metadata

Each response includes:

• Grounded Response: Initial tool-augmented answer
• Refined Response: Polished final answer
• Function Calls: List of tools used
• Grounding Metadata: Web sources (if search enabled)

Error Handling

The agent gracefully handles:

• API errors
• Tool execution failures
• Network issues
• Invalid inputs

Notes

• Mock Data: Weather data is currently mocked. Integrate a real weather API (OpenWeatherMap, WeatherAPI) for production.
• Web Scraping: Basic implementation. Use BeautifulSoup for better HTML parsing.
• Security: File operations are unrestricted. Add path validation for production.

License

This project is provided as-is for educational purposes.