Chunkr is document-intelligence infrastructure that turns complex documents — PDFs, Office files, presentations and scans — into clean, LLM-ready data. Its vision models segment every page into titled sections, tables, figures and text blocks with bounding boxes, then output structured markdown, JSON or HTML. The project has about 4,100 stars on GitHub and is actively developed in 2026.
Built by Lumina AI, Chunkr attacks the unglamorous bottleneck of every RAG pipeline: parsing. A task-based API accepts a document and a strategy (fast OCR-driven or high-quality vision-driven), processes pages asynchronously, and returns semantically chunked content that preserves reading order, table structure and layout metadata — exactly what embedding models and agents need. Outputs include per-segment types and coordinates, extracted tables as structured data, and OCR text with confidence, making it suitable for financial filings, legal contracts, research papers and government forms. The core is open source under AGPL-3.0 so teams can self-host the pipeline, while the hosted Chunkr API offers managed scaling, higher throughput and support. SDKs and a straightforward REST interface make it a drop-in parsing layer for document-heavy AI products.
Chunkr’s pipeline is strategy-aware: simple born-digital documents take a fast path, while dense scans route to vision models that read layout the way a person would, and both paths converge on the same structured output schema. It supports synchronous small jobs and large asynchronous batches, stores intermediate results for debugging, and exposes usage-based controls so teams can cap parsing spend per workload. The self-hostable core uses the same models as the hosted API, so prototypes built locally behave identically when moved to the cloud.
Chunkr is Freemium: the open-source version is free to self-host, and the hosted API offers free trial credits with pay-as-you-go pricing per page beyond that.
Chunkr is for AI engineers and product teams whose RAG or extraction quality is limited by document parsing, and who need layout-aware, structured output at scale.