RAG Retriever

A Python application that loads and processes web pages, local documents, and Confluence spaces, indexing their content using embeddings, and enabling semantic search queries. Built with a modular architecture using OpenAI embeddings and Chroma vector store.

What It Does

RAG Retriever enhances your AI coding assistant (like aider, Cursor, or Windsurf) by giving it access to:

• Documentation about new technologies and features
• Your organization's architecture decisions and coding standards
• Internal APIs and tools documentation
• Confluence spaces and documentation
• Any other knowledge that isn't part of the LLM's training data

This helps prevent hallucinations and ensures your AI assistant follows your team's practices.

Watch a Short Demo Video (not all RAG Retriever features are shown)

RAG Retriever seamlessly integrating with aider, Cursor, and Windsurf to provide accurate, up-to-date information during development.

💡 Note: While our examples focus on AI coding assistants, RAG Retriever can enhance any AI-powered development environment or tool that can execute command-line applications. Use it to augment IDEs, CLI tools, or any development workflow that needs reliable, up-to-date information.

Why Do We Need Such Tools?

Modern AI coding assistants each implement their own way of loading external context from files and web sources. However, this creates several challenges:

• Knowledge remains siloed within each tool's ecosystem
• Support for different document types and sources varies widely
• Integration with enterprise knowledge bases (Confluence, Notion, etc.) is limited
• Each tool requires learning its unique context-loading mechanisms

RAG Retriever solves these challenges by:

1. Providing a unified knowledge repository that can ingest content from diverse sources
2. Offering a simple command-line interface that works with any AI tool supporting shell commands

💡 For a detailed discussion of why centralized knowledge retrieval tools are crucial for AI-driven development, see our Why RAG Retriever guide.

Prerequisites

Core Requirements

• Python 3.10-3.12 (Download from python.org)

• pipx (Install with one of these commands):

  # On MacOS
  brew install pipx
  
  # On Windows/Linux
  python -m pip install --user pipx

---

🚀 Ready to Try It? Let's Go!

Get up and running in 10 minutes! Head over to our Getting Started Guide for a quick setup that will have your AI assistant using RAG Retriever right away.

⚡ Quick install: pipx install rag-retriever

---

Optional Dependencies

The following dependencies are only required for specific advanced features:

OCR Support (Optional)

Required only for:

• Processing scanned documents
• Extracting text from images in PDFs
• Converting images to searchable text

MacOS: brew install tesseract Windows: Install Tesseract

Advanced PDF Processing (Optional)

Required only for:

• Complex PDF layouts
• Better table extraction
• Technical document processing

MacOS: brew install poppler Windows: Install Poppler

The core functionality works without these dependencies, including:

• Basic PDF text extraction
• Markdown and text file processing
• Web content crawling
• Vector storage and search

System Requirements

The application uses Playwright with Chromium for web crawling:

• Chromium browser is automatically installed during package installation
• Sufficient disk space for Chromium (~200MB)
• Internet connection for initial setup and crawling

Note: The application will automatically download and manage Chromium installation.

Installation

Install RAG Retriever as a standalone application:

pipx install rag-retriever

This will:

• Create an isolated environment for the application
• Install all required dependencies
• Install Chromium browser automatically
• Make the rag-retriever command available in your PATH

How to Upgrade

To upgrade RAG Retriever to the latest version:

pipx upgrade rag-retriever

This will:

• Upgrade the package to the latest available version
• Preserve your existing configuration and data
• Update any new dependencies automatically

After installation, initialize the configuration:

# Initialize configuration files
rag-retriever --init

This creates a configuration file at ~/.config/rag-retriever/config.yaml (Unix/Mac) or %APPDATA%\rag-retriever\config.yaml (Windows)

Setting up your API Key

Add your OpenAI API key to your configuration file:

api:
  openai_api_key: "sk-your-api-key-here"

Security Note: During installation, RAG Retriever automatically sets strict file permissions (600) on config.yaml to ensure it's only readable by you. This helps protect your API key.

Customizing Configuration

All settings are in config.yaml. For detailed information about all configuration options, best practices, and example configurations, see our Configuration Guide.

Key configuration sections include:

# Vector store settings
vector_store:
  embedding_model: "text-embedding-3-large"
  embedding_dimensions: 3072
  chunk_size: 1000
  chunk_overlap: 200

# Local document processing
document_processing:
  supported_extensions:
    - ".md"
    - ".txt"
    - ".pdf"
  pdf_settings:
    max_file_size_mb: 50
    extract_images: false
    ocr_enabled: false
    languages: ["eng"]
    strategy: "fast"
    mode: "elements"

# Search settings
search:
  default_limit: 8
  default_score_threshold: 0.3

Data Storage

The vector store database is stored at:

• Unix/Mac: ~/.local/share/rag-retriever/chromadb/
• Windows: %LOCALAPPDATA%\rag-retriever\chromadb/

This location is automatically managed by the application and should not be modified directly.

Uninstallation

To completely remove RAG Retriever:

# Remove the application and its isolated environment
pipx uninstall rag-retriever

# Remove Playwright browsers
python -m playwright uninstall chromium

# Optional: Remove configuration and data files
# Unix/Mac:
rm -rf ~/.config/rag-retriever ~/.local/share/rag-retriever
# Windows (run in PowerShell):
Remove-Item -Recurse -Force "$env:APPDATA\rag-retriever"
Remove-Item -Recurse -Force "$env:LOCALAPPDATA\rag-retriever"

Development Setup

If you want to contribute to RAG Retriever or modify the code:

# Clone the repository
git clone https://github.com/codingthefuturewithai/rag-retriever.git
cd rag-retriever

# Create and activate virtual environment
python -m venv venv
source venv/bin/activate  # Unix/Mac
venv\Scripts\activate     # Windows

# Install in editable mode
pip install -e .

# Initialize user configuration
./scripts/run-rag.sh --init  # Unix/Mac
scripts\run-rag.bat --init   # Windows

Usage Examples

Local Document Processing

# Process a single file
rag-retriever --ingest-file path/to/document.pdf

# Process all supported files in a directory
rag-retriever --ingest-directory path/to/docs/

# Enable OCR for scanned documents (update config.yaml first)
# Set in config.yaml:
# document_processing.pdf_settings.ocr_enabled: true
rag-retriever --ingest-file scanned-document.pdf

# Enable image extraction from PDFs (update config.yaml first)
# Set in config.yaml:
# document_processing.pdf_settings.extract_images: true
rag-retriever --ingest-file document-with-images.pdf

Web Content Fetching

# Basic fetch
rag-retriever --fetch https://example.com

# With depth control (default: 2)
rag-retriever --fetch https://example.com --max-depth 2

# Enable verbose output
rag-retriever --fetch https://example.com --verbose

Web Search

# Search the web using DuckDuckGo
rag-retriever --web-search "your search query"

# Control number of results
rag-retriever --web-search "your search query" --results 10

Confluence Integration

RAG Retriever can load and index content directly from your Confluence spaces. To use this feature:

1. Configure your Confluence credentials in ~/.config/rag-retriever/config.yaml:

api:
  confluence:
    url: "https://your-domain.atlassian.net" # Your Confluence instance URL
    username: "[email protected]" # Your Confluence username/email
    api_token: "your-api-token" # API token from https://id.atlassian.com/manage-profile/security/api-tokens
    space_key: null # Optional: Default space to load from
    parent_id: null # Optional: Default parent page ID
    include_attachments: false # Whether to include attachments
    limit: 50 # Max pages per request
    max_pages: 1000 # Maximum total pages to load

2. Load content from Confluence:

# Load from configured default space
rag-retriever --confluence

# Load from specific space
rag-retriever --confluence --space-key TEAM

# Load from specific parent page
rag-retriever --confluence --parent-id 123456

# Load from specific space and parent
rag-retriever --confluence --space-key TEAM --parent-id 123456

The loaded content will be:

• Converted to markdown format
• Split into appropriate chunks
• Embedded and stored in your vector store
• Available for semantic search just like any other content

Searching Content

# Basic search
rag-retriever --query "How do I configure logging?"

# Limit results
rag-retriever --query "deployment steps" --limit 5

# Set minimum relevance score
rag-retriever --query "error handling" --score-threshold 0.7

# Get full content (default) or truncated
rag-retriever --query "database setup" --truncate

# Output in JSON format
rag-retriever --query "API endpoints" --json

Configuration Options

The configuration file (config.yaml) is organized into several sections:

Vector Store Settings

vector_store:
  persist_directory: null # Set automatically to OS-specific path
  embedding_model: "text-embedding-3-large"
  embedding_dimensions: 3072
  chunk_size: 1000 # Size of text chunks for indexing
  chunk_overlap: 200 # Overlap between chunks

Document Processing Settings

document_processing:
  # Supported file extensions
  supported_extensions:
    - ".md"
    - ".txt"
    - ".pdf"

  # Patterns to exclude from processing
  excluded_patterns:
    - ".*"
    - "node_modules/**"
    - "__pycache__/**"
    - "*.pyc"
    - ".git/**"

  # Fallback encodings for text files
  encoding_fallbacks:
    - "utf-8"
    - "latin-1"
    - "cp1252"

  # PDF processing settings
  pdf_settings:
    max_file_size_mb: 50
    extract_images: false
    ocr_enabled: false
    languages: ["eng"]
    password: null
    strategy: "fast" # Options: fast, accurate
    mode: "elements" # Options: single_page, paged, elements

Content Processing Settings

content:
  chunk_size: 2000
  chunk_overlap: 400
  # Text splitting separators (in order of preference)
  separators:
    - "\n## " # h2 headers (strongest break)
    - "\n### " # h3 headers
    - "\n#### " # h4 headers
    - "\n- " # bullet points
    - "\n• " # alternative bullet points
    - "\n\n" # paragraphs
    - ". " # sentences (weakest break)

Search Settings

search:
  default_limit: 8 # Default number of results
  default_score_threshold: 0.3 # Minimum relevance score

Browser Settings (Web Crawling)

browser:
  wait_time: 2 # Base wait time in seconds
  viewport:
    width: 1920
    height: 1080
  delays:
    before_request: [1, 3] # Min and max seconds
    after_load: [2, 4]
    after_dynamic: [1, 2]
  launch_options:
    headless: true
    channel: "chrome"
  context_options:
    bypass_csp: true
    java_script_enabled: true

Understanding Search Results

Search results include relevance scores based on cosine similarity:

• Scores range from 0 to 1, where 1 indicates perfect similarity
• Default threshold is 0.3 (configurable via search.default_score_threshold)
• Typical interpretation:
  • 0.7+: Very high relevance (nearly exact matches)
  • 0.6 - 0.7: High relevance
  • 0.5 - 0.6: Good relevance
  • 0.3 - 0.5: Moderate relevance
  • Below 0.3: Lower relevance

Features

Core Features (No Additional Dependencies)

• Web crawling and content extraction
• Basic PDF text extraction
• Markdown and text file processing
• Vector storage and semantic search
• Configuration management
• Basic document chunking and processing

Advanced Features (Optional Dependencies Required)

• OCR Processing (Requires Tesseract):

  • Scanned document processing
  • Image text extraction
  • PDF image text extraction

• Enhanced PDF Processing (Requires Poppler):

  • Complex layout handling
  • Table extraction
  • Technical document processing
  • Better handling of multi-column layouts

All core features work without installing optional dependencies. Install optional dependencies only if you need their specific features.

For more detailed usage instructions and examples, please refer to the local-document-loading.md documentation.

Project Structure

rag-retriever/
├── rag_retriever/         # Main package directory
│   ├── config/           # Configuration settings
│   ├── crawling/         # Web crawling functionality
│   ├── vectorstore/      # Vector storage operations
│   ├── search/          # Search functionality
│   └── utils/           # Utility functions

Dependencies

Key dependencies include:

• openai: For embeddings generation (text-embedding-3-large model)
• chromadb: Vector store implementation with cosine similarity
• selenium: JavaScript content rendering
• beautifulsoup4: HTML parsing
• python-dotenv: Environment management

Notes

• Uses OpenAI's text-embedding-3-large model for generating embeddings by default
• Content is automatically cleaned and structured during indexing
• Implements URL depth-based crawling control
• Vector store persists between runs unless explicitly deleted
• Uses cosine similarity for more intuitive relevance scoring
• Minimal output by default with --verbose flag for troubleshooting
• Full content display by default with --truncate option for brevity
• ⚠️ Changing chunk size/overlap settings after ingesting content may lead to inconsistent search results. Consider reprocessing existing content if these settings must be changed.

Future Development

RAG Retriever is under active development with many planned improvements. We maintain a detailed roadmap of future enhancements in our Future Features document, which outlines:

• Document lifecycle management improvements
• Integration with popular documentation platforms
• Vector store analysis and visualization
• Search quality enhancements
• Performance optimizations

While the current version is fully functional for core use cases, there are currently some limitations that will be addressed in future releases. Check the future features document for details on potential upcoming improvements.

Contributing

Please read CONTRIBUTING.md for details on our code of conduct and the process for submitting pull requests.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Command Line Options

Core options:

• --init: Initialize user configuration files
• --fetch URL: Fetch and index web content
• --max-depth N: Maximum depth for recursive URL loading (default: 2)
• --query STRING: Search query to find relevant content
• --limit N: Maximum number of results to return
• --score-threshold N: Minimum relevance score threshold
• --truncate: Truncate content in search results
• --json: Output results in JSON format
• --clean: Clean (delete) the vector store
• --verbose: Enable verbose output for troubleshooting
• --ingest-file PATH: Ingest a local file
• --ingest-directory PATH: Ingest a directory of files
• --web-search STRING: Perform DuckDuckGo web search
• --results N: Number of web search results (default: 5)
• --confluence: Load from Confluence
• --space-key STRING: Confluence space key
• --parent-id STRING: Confluence parent page ID

Web search (using DuckDuckGo)

rag-retriever --web-search "your search query" --results 5

You can then fetch content from the web search results using --fetch

rag-retriever --fetch https://found-url-from-search.com --max-depth 0

For example, to learn about new Java features:

rag-retriever --web-search "Java 23 new features guide" --results 3 rag-retriever --fetch https://www.happycoders.eu/java/java-23-features --max-depth 0

Load from Confluence (requires configuration in ~/.config/rag-retriever/config.yaml)

rag-retriever --confluence --space-key TEAM

Clean up vector store if needed

rag-retriever --clean