Question 1

How accurate is YaiGlobal Parser on Arabic OCR?

Accepted Answer

On the KITAB-Bench image-to-text track, across 3,760 samples, YaiGlobal Parser recorded a 12.5 percent character error rate and a 25.0 percent word error rate at a chrF of 85.23. All three are the best of the 17 systems compared. Results are reported per task: the parser leads on ten of thirteen headline measures across four tasks, and three measures favour a competitor.

Question 2

Is YaiGlobal's character error rate 0.125 percent?

Accepted Answer

No. The character error rate is 12.5 percent, which corresponds to a ratio of 0.125. The form 0.125 percent has circulated in error and overstates accuracy by a factor of 100. Earlier figures of 5,724 samples, 16 systems and 12 datasets are also superseded: the evaluation covers 5,949 samples across four tasks against 17 baselines, and 12 is the number of subsets within the image-to-text task alone.

Question 3

Which system performs best on KITAB-Bench?

Accepted Answer

It depends on the task; KITAB-Bench has no single overall score. In a vendor-conducted evaluation on four of its nine tasks, YaiGlobal Parser recorded the best result of 17 systems on image-to-text across character error rate, word error rate and chrF; the best of 14 on table structure at 97.99 TEDS; and the best of 11 on end-to-end PDF-to-Markdown at 79.97 MARS. The KITAB-Bench paper separately reports Gemini-2.0-Flash as the strongest system on PDF-to-Markdown among the models its authors evaluated.

Question 4

How does YaiGlobal Parser perform on Arabic handwriting?

Accepted Answer

This is where the margin is widest. Macro-averaged over the KHATT, ADAB, Muharaf and OnlineKhatt handwritten subsets, YaiGlobal Parser recorded a 14.5 percent character error rate against 22.0 percent for the strongest baseline, Gemini 2.0 Flash - a 34.09 percent relative reduction. The word error rate was 37.25 percent against 47.25 percent for AIN-7B. The parser's relative position improves as material becomes harder, whereas most baselines degrade.

Question 5

Where does YaiGlobal Parser not lead?

Accepted Answer

Three of the thirteen headline measures favour a competitor. On end-to-end text fidelity, chrF of 79.81 trails LlamaParse at 82.51, a 3.27 percent difference that most likely reflects LlamaParse's handling of embedded PDF text layers. On layout localisation, mAP at 0.5 of 0.611 trails RT-DETR in Docling at 0.750; RT-DETR is a specialised detector, and the parser accepts looser boxes in exchange for a precision of 0.916 against RT-DETR's 0.626. On layout recall, 0.711 trails RT-DETR's 0.725, a shortfall dominated by caption detection.

Question 6

Was the evaluation independently verified?

Accepted Answer

No. It was run and reported by YaiGlobal against KITAB-Bench, which is itself a third-party, peer-reviewed benchmark published by Heakl et al. in Findings of ACL 2025. It is not an independent audit. Most baseline figures are quoted from the published KITAB-Bench paper, but LlamaParse and Landing AI do not appear there and were measured internally - and LlamaParse is the strongest baseline on three of the four tasks, so those comparisons carry less weight than image-to-text, which is drawn entirely against published figures.

Question 7

Why does YaiGlobal evaluate four of the nine KITAB-Bench tasks?

Accepted Answer

The five excluded tasks are excluded on capability grounds. Chart-to-DataFrame, diagram-to-JSON and visual question answering measure semantic reasoning over a document rather than faithful extraction of its contents. Line detection and line recognition measure an internal intermediate representation that the parser does not expose as a published output. The four evaluated tasks cover 5,949 samples, which is 93.7 percent of the samples belonging to those tasks and 67.5 percent of the full benchmark.

Question 8

Can YaiGlobal Parser be deployed on-premises?

Accepted Answer

Yes. YaiGlobal offers a hosted web platform and REST API with no infrastructure overhead, and an on-premises deployment option for organisations with data sovereignty or regulatory requirements.

القسم	الصفحات	الحالة
الفصل الأول	١–٤٨	مكتمل
الفصل الثاني	٤٩–١١٢	قيد المراجعة
الملاحق	١١٣–١٣٠	مكتمل

We didn’t claim Arabic OCR leadership. We measured it.

Arabic depth you can measure — not just a claim.

Four KITAB-Bench tasks, not just OCR accuracy.

From a scanned page, out of the dark, into structured knowledge.

حفظ التراث المخطوط العربي

Arabic isn’t an afterthought at YaiGlobal.

Arabic-native intelligence

Visual grounding

Structural intelligence

Two offices, one product

The parser, packaged for the work you actually do.

Yai4Edu

Yai4Institution

Yai4Ottoman

Enterprise-ready, whichever model you choose.

Managed Cloud

Hybrid Cloud

Private & On-Premises

Built for organizations with serious archives.

Digital library creation

Search & discovery

Historical & institutional archives

Custom integration

Have Arabic documents?
Let’s benchmark them.