A local RAG (Retrieval-Augmented Generation) system that creates embeddings from your notes and documents, enabling semantic search and retrieval using Ollama and SQLite.

• Local-first: All processing happens on your machine using Ollama
• Multiple formats: Supports PDF, TXT, and Markdown files
• Semantic search: Find relevant content using natural language queries
• Efficient storage: Uses SQLite with vector extensions for fast similarity search
• Flexible chunking: Configurable chunking strategies (fixed-size or semantic)
• CLI interface: Simple command-line tools for all operations

# Clone and setup
git clone <repository-url>
cd roserade

# Create virtual environment and install
uv venv
uv pip install -e "."

# Ensure Ollama is running with the embedding model
ollama pull nomic-embed-text

# Initialize the database
roserade init

# Add documents (single file or directory)
roserade add README.md
roserade add -r ./docs/          # Recursive

# Search your documents
roserade search "python async functions"
roserade search "machine learning" --limit 5

# List indexed documents
roserade list-docs

# Remove a document
roserade remove document.pdf

Roserade uses a YAML configuration file at ~/.config/roserade/config.yaml:

database:
  path: "~/.config/roserade/index.db"

ollama:
  host: "http://localhost:11434"
  embedding_model: "nomic-embed-text"

chunking:
  strategy: "semantic" # or "fixed"
  size: 512
  overlap: 50

processing:
  supported_extensions: [".pdf", ".txt", ".md"]
  max_file_size: "50MB"

Initialize a new Roserade index database.

Options:

• --db-path: Custom database location
• --config: Custom config file

Add documents to the index.

Options:

• --recursive, -r: Process directories recursively
• --force: Re-index existing documents
• --chunk-size: Override default chunk size
• --chunk-overlap: Override default overlap
• --db-path: Custom database location

Search indexed documents using semantic similarity.

Options:

• --limit, -n: Number of results (default: 10)
• --threshold: Similarity threshold (0.0-1.0)
• --format: Output format (table/json)

List all indexed documents.

Options:

• --limit: Maximum documents to show
• --format: Output format (table/json)

Remove documents from the index.

Options:

• --pattern: Remove by pattern (glob)
• --force: Skip confirmation

# Install with dev dependencies
uv pip install -e ".[dev,test]"

# Run tests
uv run pytest

# Format code
uv run black src/ tests/
uv run ruff check src/ tests/

# Type checking
uv run mypy src/

src/roserade/
├── cli/           # Command-line interface
├── core/          # Core functionality
├── models/        # Data models
└── utils/         # Utilities

1. Document Processing: Extracts text from PDF, TXT, and MD files
2. Chunking: Breaks documents into manageable chunks using semantic or fixed-size strategies
3. Embedding: Generates vector embeddings using Ollama's nomic-embed-text model
4. Storage: Stores documents, chunks, and embeddings in SQLite with vector extensions
5. Search: Uses cosine similarity to find relevant chunks based on query embeddings

Can add the following line to you .zshrc so you can run the project from anywhere. I use the rag alias, but pick whatever works for you

rag() {
    uv run --project ~/<project-directory>/roserade roserade "$@"
}

• I'm feeling lucky style open the top search in neovim or less
• cron job for re-syncing embeddings whenever they are out of re-sync
• Experiment with making it more fine grained (sections of markdown files)
• Add the text content to the database (and/or line number of section)

https://github.com/ggozad/haiku.rag