
PaddleOCR
Turns scans, photographs and PDFs into Markdown or JSON an agent can read, in a hundred languages and on a CPU
What is PaddleOCR?
Retrieval over your own documents usually fails long before the model does, at the point where a scanned table or a two-column PDF becomes an unreadable wall of text. PaddleOCR is the layer that stops that. It recognises text in more than a hundred languages, and its document pipeline preserves what the layout meant — headings stay headings, a table comes out as a table with its cell positions, formulas and seals are handled — emitting Markdown or JSON. The models are deliberately small enough to run on a CPU or at the edge, which is what makes processing a large archive affordable. The cost is the framework underneath: this is a PaddlePaddle project, and installing it means installing that.
What can you do with PaddleOCR?
- Keep the structure, not just the words — The document pipeline emits Markdown or JSON in which headings, tables, reading order and formulas survive — which is what makes the output usable for retrieval.
- Get cell-level coordinates when you need them — The structural pipeline reports where each table cell and text block sat on the page, so an extracted figure can be traced back to the spot it came from.
- Read a hundred languages without switching models — A single recognition model covers Chinese, English, Japanese and dozens of Latin-script languages, so a mixed-language archive needs no routing step.
- Run it on the hardware you have — Model sizes range from a couple of megabytes upward, and inference runs on ordinary CPUs as well as on GPUs and several accelerators.
- Handle the awkward originals — Skewed, warped, badly lit and screen-photographed pages are explicit targets, as are seals, handwriting-adjacent scripts and industrial markings.
- Feed the tools you may already run — Dify, RAGFlow and others use it as their document layer, so the output format is one those pipelines already expect.
Before you choose PaddleOCR
- It sits on PaddlePaddle, a deep-learning framework you probably do not already have. Installation is a framework install with its own version and hardware matrix, not a single small package.
- The model line-up moves quickly and the names overlap — several generations of recognition models and two different document pipelines are current at once — so pick deliberately rather than by recency.
Star history
27 Aug to 30 Oct · +3.7k
Frequently asked questions
Is PaddleOCR free for commercial use?
PaddleOCR is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.
How can PaddleOCR be deployed?
PaddleOCR is available as Runs locally / Self-hosted.