🎙️ PodBuster

AI-Powered Podcast Fact-Checking Pipeline

PodBuster is an intelligent multi-agent system that automatically extracts and verifies factual claims from podcast transcripts using Google's Agent Development Kit (ADK) and Gemini models.

🌟 Features

• Multi-Agent Architecture: Sequential pipeline with specialized agents for different tasks
• Claim Extraction: Automatically identifies verifiable factual claims from transcripts
• Real-Time Verification: Uses Google Search for grounding and fact-checking
• Structured Output: JSON schemas ensure consistent, parseable results
• Professional Reporting: Generates Markdown tables with verdicts and citations
• Opinion Filtering: Intelligently separates facts from opinions and conversational phrases

🏗️ Architecture

PodBuster uses a three-agent sequential pipeline:

┌─────────────────┐
│  Raw Transcript │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ TranscriberAgent│  Extract verifiable claims
│  (Gemini 2.5)   │  Filter opinions/questions
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ FactCheckerAgent│  Verify each claim
│  (Gemini 2.5)   │  Use Google Search
└────────┬────────┘  Provide citations
         │
         ▼
┌─────────────────┐
│ ReporterAgent   │  Generate Markdown report
│  (Gemini 2.5)   │  Format results table
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Final Report   │
└─────────────────┘

Agent Details

1. TranscriberAgent

   • Extracts explicit, verifiable factual claims
   • Filters out opinions, emotions, questions
   • Outputs: JSON array of claim strings

2. FactCheckerAgent

   • Verifies claims using Google Search
   • Assigns status: TRUE, FALSE, or INCONCLUSIVE
   • Provides summary and authoritative source citation
   • Outputs: JSON object with status, summary, citation

3. ReporterAgent

   • Compiles all verification results
   • Generates professional Markdown table
   • Outputs: Formatted accountability report

📋 Prerequisites

• Python 3.10+
• Google API Key with Gemini access
• VS Code (recommended) or Jupyter Notebook

🚀 Setup

1. Clone and Navigate

cd /path/to/podbuster

2. Create Virtual Environment

python -m venv .venv
source .venv/bin/activate  # On macOS/Linux
# or
.venv\Scripts\activate     # On Windows

3. Install Dependencies

pip install google-adk litellm python-dotenv

4. Configure Environment

Create a .env file in the project root:

GOOGLE_API_KEY=your_google_api_key_here

5. Run the Notebook

Open podbuster.ipynb in VS Code or Jupyter and run all cells.

📁 Project Structure

podbuster/
├── .env                    # API keys (not in git)
├── .gitignore             # Git ignore rules
├── .gitattributes         # Git attributes
├── README.md              # This file
├── podbuster.ipynb        # Main pipeline notebook
├── podbuster_data.py      # Mock transcript data
└── .venv/                 # Virtual environment (not in git)

🧪 Mock Data

The project includes 4 pre-built mock transcripts for testing:

• MOCK_TRANSCRIPT_TECH: AI & Technology (ChatGPT, neurons, quantum computing)
• MOCK_TRANSCRIPT_SCIENCE: Space & Astronomy (Moon, Mars, Sun facts)
• MOCK_TRANSCRIPT_HEALTH: Nutrition & Medicine (water, bones, vitamins)
• MOCK_TRANSCRIPT_HISTORY: Historical Events (Declaration, WWII, Great Wall)

Each transcript contains a mix of:

• ✅ True facts (verifiable and correct)
• ❌ False claims (verifiable but incorrect)
• 💭 Opinions (should be filtered out by TranscriberAgent)

💻 Usage

Basic Usage

# In the notebook, after running all setup cells:
response = await podbuster_runner.run_debug(MOCK_TRANSCRIPT)

Custom Transcript

from podbuster_data import MOCK_TRANSCRIPT_TECH

response = await podbuster_runner.run_debug(MOCK_TRANSCRIPT_TECH)

With f-string

custom_transcript = "Your custom podcast transcript here..."
response = await podbuster_runner.run_debug(
    f"Analyze this podcast transcript: {custom_transcript}"
)

🔧 Configuration

Retry Configuration

The pipeline includes automatic retry logic for API failures:

retry_config = types.HttpRetryOptions(
    attempts=5,
    exp_base=7,
    initial_delay=1,
    http_status_codes=[429, 500, 503, 504]
)

Model Configuration

• TranscriberAgent: gemini-2.5-flash-lite (fast claim extraction)
• FactCheckerAgent: gemini-2.5-flash-lite (thorough verification)
• ReporterAgent: gemini-2.5-flash-lite (report generation)

📊 Output Format

The pipeline generates a Markdown table with:

Claim	Status	Summary	Citation
"ChatGPT was released in November 2022"	TRUE	Verified via official OpenAI announcements	https://openai.com/...
"Quantum computers can solve any problem instantly"	FALSE	Quantum computers have limitations...	https://...

🛠️ Troubleshooting

Import Errors

Make sure all dependencies are installed:

pip install google-adk litellm python-dotenv

API Key Issues

Verify your .env file contains a valid Google API key:

cat .env  # Should show: GOOGLE_API_KEY=...

Context Variable Errors

• Use MOCK_TRANSCRIPT directly, not {MOCK_TRANSCRIPT} in strings
• {variable} syntax is for ADK context variables, not Python variables

🤝 Contributing

This is a Kaggle competition project. Feel free to fork and experiment!

📄 License

MIT

🔗 Technologies

• Google Agent Development Kit (ADK)
• Gemini API
• LiteLLM
• Python 3.10+

📝 Notes

• The pipeline uses in-memory sessions (no persistence)
• Google Search tool requires active internet connection
• Fact-checking accuracy depends on search result quality
• Results may vary based on current web information