lumina-ai-inc/chunkr

What it solves

Chunkr is a document intelligence API designed to convert complex documents—such as PDFs, PowerPoint presentations, Word documents, and images—into structured, LLM-ready chunks. It solves the problem of preparing high-quality data for Retrieval-Augmented Generation (RAG) pipelines by handling layout analysis, OCR, and semantic chunking.

How it works

The project provides a production-ready service that processes documents through a pipeline of layout analysis, OCR (with bounding boxes), and Vision-Language Model (VLM) processing. It transforms raw files into structured HTML and Markdown formats, which are then semantically chunked for use in LLMs. The open-source version utilizes community and open-source models for these tasks, while a managed Cloud API version uses proprietary models for higher accuracy.

Who it’s for

It is intended for developers and AI engineers building RAG applications who need to automate the document parsing and chunking process for various file formats.

Highlights

• Multi-format support: Processes PDFs, PPTs, Word docs, and images.
• Structured output: Generates HTML and Markdown for better LLM consumption.
• Layout Analysis: Performs layout analysis and OCR with bounding boxes.
• Flexible LLM Configuration: Supports multiple LLM providers (OpenAI, Google AI Studio, OpenRouter, and self-hosted options like vLLM or Ollama) via a models.yaml configuration file.
• Docker-ready: Easy deployment via Docker Compose with support for GPU, CPU-only, and Mac ARM architectures.