Chunkr is an open-source document intelligence API designed to transform complex documents into RAG/LLM-ready data. It targets developers and researchers needing to process PDFs, PPTs, Word docs, and images for AI applications, offering layout analysis, OCR, and semantic chunking.

How It Works

Chunkr leverages a multi-stage pipeline that includes document parsing, optical character recognition (OCR) with bounding boxes, and layout analysis to generate structured HTML and Markdown outputs. It supports Vision-Language Models (VLMs) for enhanced understanding and offers flexible LLM integration via configuration files or environment variables, enabling users to select and manage various LLM providers.

Quick Start & Requirements

• Install: pip install chunkr-ai
• Prerequisites: Docker and Docker Compose for self-hosting. NVIDIA Container Toolkit is recommended for GPU support.
• Self-Hosted Deployment: Requires cloning the repository, setting up .env and models.yaml files, and running via docker compose up -d (with variations for CPU and Mac ARM).
• Documentation: chunkr.ai

Highlighted Details

• Supports multiple document formats (PDF, PPT, Word, images).
• Provides structured output options: HTML, Markdown, plain text, JSON.
• Offers self-hosted deployment via Docker Compose and Kubernetes (Helm chart available).
• Flexible LLM configuration supporting multiple providers and rate limiting.

Maintenance & Community

• Contact: mehul@chunkr.ai
• Website: chunkr.ai
• Community: Discord link available in README.

Licensing & Compatibility

• License: Dual-licensed: GNU Affero General Public License v3.0 (AGPL-3.0) and a Commercial License.
• Compatibility: AGPL-3.0 terms may require derivative works to be open-sourced if linked. Commercial use requires a separate license.

Limitations & Caveats

The AGPL-3.0 license imposes significant obligations on users who modify or distribute the software, potentially requiring them to open-source their own code. Specific VLM processing controls are mentioned but not detailed in the README.