Frequently Asked Questions

YaiGlobal is a document-intelligence company that builds Arabic OCR: software that turns scanned Arabic pages into structured, searchable text. It was founded in 2018 and has offices in Santa Clara, California (USA) and Ariana (Tunisia), with a native Arabic-speaking team that built its engine Arabic first.

That engine, YaiGlobal Parser, does more than recognize characters. It keeps headings, tables and reading order, and traces every extracted element back to the region of the page it came from. YaiGlobal packages the parser into platforms for libraries, universities, archives and researchers, and builds custom solutions for governments, publishers, banks, legal organizations and enterprises with large Arabic document collections.

Read more about why Arabic OCR is harder than Latin-script OCR and how YaiGlobal Parser performed on KITAB-Bench.

YaiGlobal provides Arabic OCR and AI-powered document processing, delivered as three platforms, as direct access to its parser, and as custom projects.

• Yai4Edu digitizes collections of documents. It combines OCR with multilingual translation, summarization, extraction of tables, graphs and charts, text-to-speech, smart bookmarking and Arabic diacritization (تشكيل), and exports to Word, searchable PDF or EPUB.
• Yai4Institution turns the collection of a library, university or archive into a digital archive. It adds role-based access for researchers, librarians and administrators, a searchable institutional repository, activity reporting, and HTML export alongside Word, searchable PDF and EPUB.
• Yai4Ottoman processes Ottoman Turkish manuscripts: OCR, translation into English or Arabic with historical context, and extraction of metadata and named entities for indexing.
• YaiGlobal Parser is available through a hosted web platform and REST API, in a hybrid cloud setup, or deployed on your own hardware or private cloud.
• Custom solutions cover digital library creation, search and discovery (full-text search, citation linking and retrieval-augmented question answering over a collection), historical and institutional archives, and integration with existing systems.

All three platforms include a human-in-the-loop editor, so people who know the material can review and correct the output. See deployment options and custom solutions on the homepage.

YaiGlobal processes Arabic documents, printed and handwritten, modern and historical, as well as Ottoman Turkish manuscripts through Yai4Ottoman. Its engine was developed and tested on the features that make Arabic difficult: connected letterforms, letters that change shape by position, optional diacritics, calligraphy and degraded historical print.

Examples of the material it handles:

• Printed Arabic text, including degraded historical print.
• Handwritten Arabic, including manuscripts. Its published handwriting results cover four kinds of material: paragraphs and lines, isolated words, historical manuscripts, and handwriting captured as pen strokes.
• Documents with tables, captions, footnotes and multi-region layouts, where structure matters as much as the text.
• Pages that mix Arabic with Latin text and Western or Arabic-Indic digits, where reading order has to be preserved.
• PDFs, converted end to end into structured Markdown.
• Scanned Ottoman Turkish documents and manuscripts, through Yai4Ottoman.

Accuracy depends on the material more than on the system, and the benchmark behind YaiGlobal's published results under-represents some document types, such as governmental and financial records. The reliable way to know how it performs on yours is to send a representative sample: YaiGlobal will run it and return the output with error rates measured on your material. For handwritten archives, see Arabic handwriting recognition.

YaiGlobal Parser is YaiGlobal's Arabic OCR and document-parsing engine, and the technology behind Yai4Edu, Yai4Institution and Yai4Ottoman. It turns a scanned page into structured output rather than a flat stream of characters:

• Text in correct reading order, including bidirectional lines that mix Arabic and Latin script.
• Tables with their row, column and span structure intact.
• Layout decomposed into typed regions such as headings, body text, tables and captions.
• Region-level provenance: every extracted element is traced back to the area of the page it came from, so the output can be checked against the source.

It recognizes printed text, handwriting and degraded historical material, and converts PDFs end to end into Markdown. It is available through YaiGlobal's hosted web platform and REST API, in a hybrid setup, or on premises for organizations with data-sovereignty or regulatory requirements. In YaiGlobal's own evaluation on KITAB-Bench, it recorded the lowest character error rate of the 17 systems compared on Arabic image-to-text; see the main results and the benchmark page.

KITAB-Bench is a peer-reviewed benchmark for Arabic OCR and document understanding, published by Heakl and colleagues in the Findings of ACL 2025. It contains 8,809 samples across nine tasks and 36 sub-domains, covering printed text, handwriting, tables, charts, diagrams, layout analysis and end-to-end PDF conversion.

KITAB-Bench is not YaiGlobal's benchmark. It was built and peer-reviewed by a research group unaffiliated with YaiGlobal, and YaiGlobal did not design it, contribute to it or influence its composition. Its datasets, task definitions, metrics and evaluation code are the work of its authors. YaiGlobal ran its own parser on the benchmark and reports the results against the published figures for other systems.

Sources: the paper in the ACL Anthology and on arXiv (2502.14949). YaiGlobal's results: Arabic OCR benchmark results on KITAB-Bench.

Four of the nine KITAB-Bench tasks: image-to-text recognition, table structure recognition, end-to-end PDF-to-Markdown conversion, and layout detection.

• Image-to-text recognition measures how accurately text is transcribed from a page image, using character error rate (CER), word error rate (WER) and chrF. It includes four handwritten corpora.
• Table structure recognition measures whether a table's rows, columns and cells are reconstructed correctly, using TEDS.
• PDF-to-Markdown measures the conversion of a whole PDF into structured Markdown, where recognition, reading-order and structure errors all add up, using MARS.
• Layout detection measures how well a system finds and labels the regions of a page, using precision, recall, F1 and mAP.

The other five tasks were not evaluated. Chart-to-DataFrame, diagram-to-JSON and visual question answering measure reasoning about a document rather than faithful extraction of its contents. Line detection and line recognition measure an intermediate step the parser does not expose as output. No result from this evaluation applies to those five tasks. The evaluation methodology explains each decision, and the glossary of OCR metrics defines every measure.

5,949 samples across four tasks, against seventeen baselines in total. The number of systems compared differs by task, so each ranking holds only for its own task:

• Image-to-text recognition: 3,760 samples, 17 systems compared.
• Layout detection: 1,700 samples from the Arabic BCE-Arabic-v1 pool, 6 systems compared.
• Table structure recognition: 456 samples, 14 systems compared.
• PDF-to-Markdown: 33 samples, 11 systems compared.

The 5,949 samples are 67.5% of the 8,809-sample benchmark. Three tasks were scored on their complete pool. Layout detection excluded 400 English DocLayNet pages, because the subject of the study is Arabic and the published baselines report the two pools separately.

The systems compared include Gemini 2.0 Flash, GPT-4o, AIN-7B, Azure Document Intelligence OCR, LlamaParse, Tesseract, EasyOCR and PaddleOCR. Most of their figures are quoted from the published KITAB-Bench paper. LlamaParse and Landing AI do not appear in that paper and were measured by YaiGlobal. Every system and measure is listed in the full KITAB-Bench results.

YaiGlobal Parser led on ten of thirteen headline measures. On the other three, a baseline led, and those results are published too.

• Image-to-text: a character error rate of 12.50%, a word error rate of 25.00% and a chrF of 85.23, the best of the 17 systems compared on all three. The next-lowest character error rate was Gemini 2.0 Flash at 13.00%.
• Handwriting: a character error rate of 14.50% against 22.00% for Gemini 2.0 Flash, a 34.09% relative reduction. The margin is wider on handwriting than on the full task.
• Tables: TEDS 97.99, the highest of 14 systems, against 90.55 for LlamaParse. All thirteen table types scored above 93.
• PDF-to-Markdown: MARS 79.97, the highest of 11 systems. Text fidelity (chrF 79.81) trails LlamaParse at 82.51.
• Layout detection: precision 0.916 and F1 0.801, both the highest. Recall (0.711) and mAP@0.5 (0.611) trail DETR in Docling, at 0.725 and 0.750.

Read these figures with their limits. The evaluation was conducted by YaiGlobal, not by an independent party, and has not been replicated. The LlamaParse and Landing AI figures were measured internally. Image-to-text averages follow the KITAB-Bench convention of an unweighted mean across twelve subsets; weighted by sample count, the character error rate is 14.08%. Error rates are percentages: 12.50% is a ratio of 0.125, not 0.125%. Details are on the benchmark results page and the methodology page, which lists six limitations.

Yes. In Yai4Edu and Yai4Institution, you choose the services a document needs, and each one becomes a stage in that document's processing pipeline, with its own status. A single document can go through OCR, Markdown conversion, diacritization, translation, summarization, bookmarking and text-to-speech in one pipeline.

A human-in-the-loop editor then lets you review and correct the output of each stage, such as the OCR text, the translation and the diacritized Arabic, side by side with the source page. Yai4Institution adds custom workflows that follow institutional policies and automate OCR, translation and compliance checks. Yai4Ottoman runs the same kind of pipeline for Ottoman Turkish: OCR, translation into English or Arabic, then extraction of metadata and named entities.

To build a workflow into your own systems, YaiGlobal offers REST API access and custom integration tuned to your workflow. Product details: Yai4Edu, Yai4Institution and Yai4Ottoman.

YaiGlobal is built for organizations that digitize whole collections, not single pages, and its deployment options include managed capacity for large processing runs. Its platforms take a scanned collection through to a structured, searchable digital library.

• Managed cloud: the hosted platform and REST API, with updates, maintenance and scaling managed by YaiGlobal.
• Hybrid cloud: you decide for each collection where its documents are processed. Sensitive material stays in your environment, and managed cloud capacity handles large processing runs.
• Private and on-premises: the engine runs entirely on your own hardware or private cloud.

Once processed, a collection is held in a digital database with full-text search, pagination, and filters by subject and language, so a large archive stays easy to navigate. To plan a project for your collection, contact YaiGlobal with a description of the material and a representative sample.