Self-hosted document parser that admits the cloud is better

What it does

Chunkr is a self-hosted document intelligence API written in Rust. It ingests PDFs, PowerPoints, Word documents, and images, then performs layout analysis, OCR, and vision-language model processing to emit structured HTML, Markdown, and semantically chunked output meant for retrieval-augmented generation pipelines. It exposes a web UI and an API, and runs either GPU-accelerated or CPU-only.

The interesting bit

The README is unusually candid about the open-source version’s place in the world: it uses community models and is recommended for development and testing, while the company’s managed Cloud API runs proprietary in-house models for higher accuracy and speed. That transparency makes the repo a credible on-ramp rather than a bait-and-switch.

Key highlights

• Produces structured HTML and Markdown alongside OCR bounding boxes.
• Supports multiple LLM backends via models.yaml or environment variables, including OpenAI, Google AI Studio, OpenRouter, or self-hosted options like VLLM and Ollama.
• Runs fully offline and self-hosted with optional GPU acceleration.
• Dual-licensed under AGPL-3.0 with a commercial license available.

Caveats

• The open-source version lacks Excel support, which is reserved for the Cloud and Enterprise tiers.
• The README explicitly recommends the proprietary Cloud API for production workloads, positioning the open-source release as “good” rather than “best” for accuracy and reliability.
• AGPL-3.0 licensing may complicate integration into proprietary SaaS products without a commercial license.

Verdict

A solid choice if you need a transparent, self-hosted document preprocessing pipeline and are comfortable tuning open-source models. Look elsewhere if you require native Excel parsing, guaranteed production-grade accuracy out of the box, or cannot accommodate copyleft licensing.