PDF Processor for Embeddings

Why selective OCR

OCR is the slowest, most expensive and most error-prone step in a document pipeline. Running it on every page means paying for — and risking invented text on — pages whose text layer was already fine. PDF Classifier decides page by page; Chandra OCR 2 only ever sees the pages it is told to read, and every page still ends up in the output, read by one route or the other.

What it does

Batch in, decisions first. Add PDFs and images, preview each routing decision before anything runs, then watch per-page progress.
Originals are left alone. A PDF with no classification yet is classified on a temporary copy.
Doubt stops the line. A document the classifier marks for manual review is held back rather than read end to end by OCR. A manual page range, and a launcher for PDF Classifier’s own page review, are there when you need them.
Choice of engine. Chandra runs on a local GPU or through a vLLM server; MinerU and a native-text-only mode that needs no GPU are alternatives.
Output for retrieval. Canonical JSON, embedding-ready JSONL and a QC report per document, optionally Markdown and HTML, in a folder tree that mirrors the input.
Language per element. English, Russian and Azerbaijani are detected paragraph by paragraph, not once per document.

Nothing is overwritten

Instead of

Replacing the last run’s output with this one’s. Every processing pass is written as a new revision, staged and moved into place only when complete. Earlier revisions are kept up to a retention count you set, and removing a document from the list never removes it from disk.

Status, stated plainly

Under active development and run from source on Windows; the installer has not been built yet, and macOS is unverified. Chandra OCR needs an NVIDIA GPU with at least 12 GiB of free memory, and the application refuses to start it — and says why — when that is not available. Reading order on multi-column native pages is not yet reliable.

The vLLM server address can be changed: pointed at a machine other than your own, page images are sent to it.

Licensing

The application is licensed under Apache 2.0 and includes the Chandra inference code, also Apache 2.0, unchanged. The Chandra OCR 2 model weights are not included and carry their own licence, a modified OpenRAIL-M. It does not permit use by an organisation with more than US$2 million in annual revenue or in funding raised, or by one offering a product that competes with the model’s publisher, Datalab; personal and research use are exempt. It also asks for attribution and extends its terms to the model’s output. Check those terms before using the model for work.