Chunky

A native macOS application for processing documents into semantic chunks and storing them as vector embeddings in a Qdrant vector database.

Features

• Multi-format document processing (PDF, DOCX, PPTX, HTML, images) via Docling
• Local VLM support with DocTags output (SmolDocling MLX, Granite-Docling MLX)
• Text chunking with configurable token limits and overlap
• Embedding generation via Ollama (local LLM server)
• Vector storage in Qdrant via gRPC
• Collection browser for exploring stored vectors
• Export to Markdown, HTML, JSON, and plain text

Prerequisites

• Xcode
• Docker (for Qdrant)
• Ollama
• Python 3 (for Docling)

brew install protobuf swift-protobuf grpc-swift

Setup

make first-launch
make setup-venv

The Python venv is created at ~/.chunky-venv with Docling and MLX installed.

Note: The build disables code signing by default, so no Apple Developer account is required.

Usage

make run           # build and launch the app
make build         # build only
make test          # run Swift tests
make test-python   # run Python tests (no docling needed)
make test-all      # run all tests
make clean         # clean build artifacts
make open-xcode    # open in Xcode

Start the required services:

make qdrant        # start Qdrant on localhost:6333 (REST) / 6334 (gRPC)
make ollama        # start Ollama
make services      # start both

VLM Models

Chunky supports local vision-language models for document processing via Docling's built-in model specs. These run directly on Apple Silicon using MLX — no Ollama needed for VLM inference.

Model	Engine	Output	Speed*
SmolDocling 256M (MLX)	MLX	DocTags	~6s/page
SmolDocling 256M (Transformers)	Transformers	DocTags	~102s/page
Granite-Docling 258M (MLX)	MLX	DocTags	—
Granite-Docling 258M (Transformers)	Transformers	DocTags	—

*Benchmarked on MacBook M3 Max, single PDF page.

The default model is SmolDocling MLX. Change it in Preferences > Documents > VLM Model. Select "Standard (no VLM)" to use the classic layout-analysis pipeline instead.

Architecture

Chunky/Sources/
  Models/         -- AppState, AppSettings, ProcessingJob, QdrantModels
  Services/       -- Docling, Ollama, Qdrant, Chunking, DocumentProcessor
  Views/          -- ContentView, CollectionBrowser, Preferences
  Theme/          -- AppTheme
LocalPackages/    -- qdrant-swift (local gRPC client package)

Dependencies

• grpc-swift-2 -- GRPCCore
• grpc-swift-nio-transport -- GRPCNIOTransportHTTP2
• grpc-swift-protobuf -- GRPCProtobuf
• swift-protobuf
• swift-collections

Resources

• Docling