DoppelCheck is a sophisticated web content analysis tool designed to help users critically evaluate online information. It supports informed decision-making by providing automated fact-checking and information verification capabilities through a browser bookmarklet.

You can find extensive documentation here.

DoppelCheck adds a sidebar to web pages that allows users to:

1. Extract key points from web page content
2. Find supporting evidence for claims from external sources
3. Compare and verify information between original content and external sources
4. Highlight relevant passages in the original document
5. Customize search sources for verification

The tool is designed to strengthen digital literacy and support critical thinking about online information. It runs locally for enhanced privacy and security while being highly customizable.

DoppelCheck abstracts the process of information verification through these steps:

stateDiagram-v2
    [*] --> content
    content --> essence: extract
    essence --> sources: retrieve
    sources --> content: select new
    sources --> support: compare to content
    support --> [*]

If you're new to bookmarklets, they're simply special browser bookmarks that contain JavaScript code instead of a website address. When you click a bookmarklet, it runs this code on the current webpage you're viewing, adding new functionality without requiring a browser extension.

With the DoppelCheck bookmarklet:

• It's added to your browser's bookmarks bar just like any normal bookmark
• It works across most modern browsers (Chrome, Firefox, Safari, Edge)
• No installation in the traditional sense - just add it to your bookmarks
• It only activates when you click it, giving you complete control

• Content Extraction: Automatically identifies and summarizes key points from web content
• Evidence Retrieval: Searches for supporting evidence from external sources
• Alignment Analysis: Compares original content with external sources to verify accuracy
• Visual Highlighting: Highlights relevant text on the original page
• Customizable Sources: Add specific domains as trusted search sources
• Local Processing: Runs locally on your machine for privacy
• Selection Summary: Summarize any selected text with one click
• Multilingual Support: Works with German and English content

DoppelCheck can be customized by editing the content_analysis_config.json file:

{
    "host": "0.0.0.0",
    "port": 5000,
    "model": "tulu3",
    "ollama_host": "https://localhost:11434",
    "debug": false,
    "custom_urls": [
        "https://sueddeutsche.de",
        "https://tagesschau.de"
    ],
    "google_custom_search_api_key": "YOUR_API_KEY_HERE",
    "google_custom_search_engine_id": "YOUR_SEARCH_ENGINE_ID_HERE"
}

Settings you can change:

• host: The server address (usually leave as is)
• port: The network port to use (change if port 5000 is already in use)
• model: The Ollama AI model to use (tulu3 is recommended)
• ollama_host: The address of your Ollama installation
• debug: Set to "carview.php?tsp=true" if you're experiencing issues
• custom_urls: Pre-configured trusted domains for searches
• google_custom_search_api_key: Your Google API key (replace with your actual key)
• google_custom_search_engine_id: Your Google Search Engine ID (replace with your actual ID)

If you want to use your own Google search capabilities instead of the default ones, you'll need to obtain:

• Google Custom Search API Key: Get it from Google Cloud Console
• Google Custom Search Engine ID: Create a new search engine and get the ID from Programmable Search Engine Control Panel

Then update these values in your content_analysis_config.json file. The example values shown in the configuration above are placeholders and will not work - you must replace them with your own credentials.

Note: When running with Docker, the Ollama service runs on port 11435 (mapped from container port 11434), but DoppelCheck is configured to connect to it internally on port 11434. If you're running Ollama separately outside of Docker, use port 11434 in your configuration.

DoppelCheck requires two main components to work properly:

1. The local DoppelCheck server
2. The DoppelCheck bookmarklet in your browser

Follow these steps in order to set up DoppelCheck:

Before you begin, make sure you have the following installed:

• Docker and Docker Compose
• For GPU acceleration (optional):
  • NVIDIA GPU: NVIDIA Container Toolkit
  • AMD GPU: ROCm drivers installed on host

1. Clone the repository:

   git clone https://github.com/Doppelcheck/main
   cd main

Run DoppelCheck with one of the following commands:

# CPU-only mode (default)
docker compose up
# With NVIDIA GPU acceleration
sudo docker compose --profile nvidia up
# With AMD ROCm acceleration
sudo docker compose --profile amd up

Wait for the startup process to complete. You should see log messages indicating that both the DoppelCheck server and the Ollama LLM service are running. The DoppelCheck server will be available at https://localhost:5000.

Once the server is running:

1. Open your browser and navigate to https://localhost:5000
2. Make sure your bookmarks bar is visible (press Ctrl+Shift+B in Chrome/Firefox on Windows/Linux or Command+Shift+B on Mac)
3. Find the "doppelcheck⠆" button on the DoppelCheck homepage
4. Drag this button to your bookmarks bar
5. The bookmarklet is now installed and ready to use on any webpage

sequenceDiagram
participant Client
participant Server
note over Client,Server: User drags doppelcheck link to bookmarks
Server-->>Client: bookmarklet javascript code
note over Client: browser displays "doppelcheck" icon in bookmarks

Now you can use DoppelCheck on any webpage:

1. Navigate to a webpage you want to analyze
2. Click the "doppelcheck⠆" bookmark in your bookmarks bar
3. DoppelCheck will analyze the page and display results in a sidebar

Important: Remember that both the Docker containers must be running whenever you want to use DoppelCheck. If you restart your computer or stop the Docker containers, you'll need to start the DoppelCheck server again (step 3) before using the bookmarklet.

If you're planning to use GPU acceleration with an NVIDIA GPU, you need to properly install and configure the NVIDIA Container Toolkit:

1. Install the NVIDIA Container Toolkit:

   # Add the NVIDIA Container Toolkit repository
   curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey | sudo apt-key add -
   distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
   curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list | sudo tee /etc/apt/sources.list.d/nvidia-container-runtime.list
   sudo apt-get update
   # Install nvidia-container-toolkit package
   sudo apt-get install -y nvidia-container-toolkit

2. Configure Docker to use the NVIDIA runtime:

   sudo nvidia-ctk runtime configure --runtime=docker

3. Restart Docker to apply the changes:

   sudo systemctl restart docker

4. Verify your NVIDIA driver installation:

   nvidia-smi

   This command should display information about your GPU and NVIDIA driver version. If it fails, you may need to install or update your NVIDIA drivers first.

5. Verify the NVIDIA Container Toolkit setup:

   sudo docker run --rm --gpus all pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime nvidia-smi

   Using the same image as in your Dockerfile (pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime) ensures that you're testing with the same CUDA version (12.1) that your application will use. This is more relevant than using a generic CUDA image that might have a different version.

When running DoppelCheck with Docker, you might encounter package dependency conflicts, especially between PyTorch, torchvision, and transformers packages. This typically happens because the CUDA-compiled version of certain operators isn't compatible with your current setup.

If you encounter errors like RuntimeError: operator torchvision::nms does not exist or NumPy compatibility issues, you can try:

1. Fix in Docker environment:

   # Access the running container
   docker exec -it doppelcheck-nvidia bash
   # Inside the container, install compatible versions
   pip install torch==2.1.2 torchvision==0.16.2 transformers==4.41.0 --extra-index-url https://download.pytorch.org/whl/cu121
   pip install numpy<2.0.0

2. Update requirements.txt: If rebuilding the image, modify the requirements.txt file to use compatible versions:

   torch==2.1.2
   torchvision==0.16.2
   transformers==4.41.0
   --extra-index-url https://download.pytorch.org/whl/cu121
   numpy<2.0.0
   

3. Force CPU mode: In some cases, forcing CPU mode can bypass GPU compatibility issues:

   # Run in CPU-only mode
   docker compose up

When you use DoppelCheck, it follows this workflow:

sequenceDiagram
    participant Client
    participant Server
    participant LLM_Interface as LLM Interface
    note over Client, Server: User clicks "doppelcheck" icon
    Client->>Server: website body
    note over Client: add sidebar element to current website DOM
    Server->>LLM_Interface: request key claims
    note right of Server: waiting for LLM interface
    LLM_Interface->>Server: key claims
    loop for each CLAIM
        Server->>Client: CLAIM
        note over Client: display CLAIM in sidebar
        note over Client: add "check alignment" button
    end

When you request to check the alignment of content with external sources:

sequenceDiagram
    participant Client
    participant Server
    participant Data_Source as Data Source
    participant LLM_Interface as LLM Interface
    note over Client,Server: User clicks "check alignment" button
    Client->>Server: CLAIM
    Server->>Data_Source: request documents relevant to claim
    note right of Server: waiting for data source interface
    Data_Source->>Server: relevant documents
    loop for each DOCUMENT
        Server->>Client: each DOCUMENT uri
        note over Client: display uri
        Server->>LLM_Interface: request match between each DOCUMENT and CLAIM
        note right of Data_Source: waiting for LLM interface
        LLM_Interface->>Server: MATCH
        Server->>Client: MATCH
        note over Client: update source URI with MATCH
    end

• Check if Python is correctly installed: Run python --version or python3 --version
• Ensure Ollama is running: Open a new terminal and run ollama list
• Check if port 5000 is already in use: Try changing the port in the config file

• Make sure the DoppelCheck server is running
• Check if your browser is blocking JavaScript
• Try using a different browser
• Refresh the page and try again

• The page might not contain suitable text content for analysis
• Try a different webpage with more text content
• Check if the website blocks content extraction

• Check your internet connection
• Try different search terms
• Add more specific trusted domains

• Make sure you've installed Ollama from ollama.ai
• Restart your computer
• Check if Ollama is running in the background

If you continue to experience issues:

• Check the GitHub repository for known issues
• Submit a bug report with detailed information about your problem
• Contact support at info@doppelcheck.com

Q: Is my data private when using DoppelCheck?
A: Yes! DoppelCheck processes everything locally on your computer. The content of pages you analyze never leaves your machine except when searching for supporting evidence.

Q: Do I need an internet connection?
A: Yes, for two reasons: to browse webpages and to search for supporting evidence. The core analysis happens locally.

Q: Can I use DoppelCheck in languages other than English?
A: Yes, DoppelCheck works well with both English and German content.

Q: How accurate is the analysis?
A: DoppelCheck uses AI to identify key points and compare sources. The accuracy depends on the quality and availability of related content online. Always use your judgment when evaluating results.

Q: Will DoppelCheck slow down my browsing?
A: The bookmarklet only activates when you click it, so it won't affect normal browsing. When analyzing a page, it uses some computer resources, but modern computers can handle this without issue.

Q: Is DoppelCheck free to use?
A: Yes, DoppelCheck is a free, open-source project supported by WPK-Innovationsfonds and Media Lab Bayern.

For those interested in the technical details:

• server/: Main server code
  • app.py: FastAPI application
  • bookmarklet.js: Client-side JavaScript code
  • routes.py: API routes
  • models.py: Data models
  • chunkers/: Text chunking algorithms
  • retrievers/: Search and retrieval modules
  • services/: Core services for extraction, comparison, etc.
  • templates/: HTML templates
  • static/: Static assets (CSS, JS, images)
  • utils/: Utility functions

For development, set up a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

Start the server in debug mode:

python main.py --debug

DoppelCheck is an ongoing project. Contributions are welcome through:

• Bug reports and feature requests: If you find issues or have ideas for improvements, please open an issue in our GitHub repository.
• Code contributions: Pull requests are highly appreciated. Please ensure your code adheres to the project's coding standards and include proper documentation.
• Documentation improvements: Help us make DoppelCheck more accessible by improving our documentation.
• Testing and feedback: Try DoppelCheck in different environments and let us know how it works.
• Plugin Development: Extend DoppelCheck's functionality by developing plugins for additional data sources or enhanced capabilities.

When submitting pull requests, please:

1. Focus on a single feature or bug fix per PR
2. Include appropriate tests
3. Update documentation if needed
4. Follow existing code styles

This project uses components with various licenses:

• Code: MIT License
• IBM Plex fonts: SIL Open Font License 1.1

DoppelCheck is a project by AI scientist and consultant Mark Wernsdorfer, supported by funding programs from WPK-Innovationsfonds and Media Lab Bayern. It aims to promote critical evaluation of online information and strengthen digital literacy through technology.

Special thanks to our contributors and supporters for their invaluable assistance in this project:

• Website: Mark Wernsdorfer
• Project Links:
  • WPK Innovationsfonds
  • Media Lab Bayern
  • GitHub