Document Intelligence
Arabic paperwork, read structurally, not just scanned
A scanned Arabic archive is not a searchable archive.
Physical Arabic documents are a dead end for most AI pipelines. Cursive script, diacritics, right-to-left layout and complex forms trip up general-purpose OCR, leaving scanned government archives, contracts and manuscripts unsearchable.
Delivered by
Baseer turns scanned archives into structured, searchable Markdown, and Baseer Extract turns that into data - invoice line items, contract clauses, form fields - an ERP or workflow can consume directly.
Why this works on Arabic when general OCR does not
Trained on the script, not adapted to it
Baseer is a 3-billion parameter vision-language model fine-tuned specifically for Arabic, on a large-scale dataset combining synthetic and real-world Arabic documents. Cursive script, diacritics and regional typographic variation are what it was built for.
Structure survives the read
Standard OCR returns a raw text string. Baseer reads columns, tables, headers and layout hierarchy, then emits clean Markdown you can index and build on.
Measured on an Arabic benchmark
Accuracy is reported on Misraj-DocOCR, an expert-verified Arabic OCR benchmark we built and published, rather than on an English test set with Arabic bolted on.
What it reads, and what it gives back
Layout understanding
Multi-column pages, embedded tables and complex formatting come back with their logical structure intact instead of scrambled into a flat block of text.
Native Arabic support
The failure modes that trip general-purpose OCR on Arabic: cursive script, diverse fonts, diacritics, right-to-left orientation and regional typographic variation.
Field extraction, not just text
Baseer Extract turns the read into data: invoice line items, contract clauses and dates, form fields and identity documents, delivered as structured output an ERP or workflow consumes directly.
Handwriting
The Nakba variant handles handwritten Arabic, so a manuscript or a hand-filled form is in scope rather than an exception.
The numbers behind it
- 0.25Word Error Rate on Misraj-DocOCR, state of the art for Arabic document OCR
- 3BParameter vision-language model, fine-tuned for Arabic
What changes
The archive becomes searchable
Scanned government records, contracts and manuscripts stop being images and start being text a search index and a RAG pipeline can actually use.
Nobody retypes a document
Structured output goes straight into the receiving system, so the manual re-keying step (and the error rate that comes with it) leaves the process.
Backlogs stop being permanent
Volume is a throughput question rather than a hiring question, which is the difference between clearing a backlog and living with it.
How it deploys
Cloud API
Through Kawn Console, for teams building RAG pipelines or document applications on top.
On-premises
For organisations with data sovereignty requirements. The documents never leave your environment.
Direct
Use it as a product at baseerocr.com without an integration project.
By Industry
Let's talk about what you're trying to build.
Tell us the problem. We'll tell you honestly whether AI is the right answer.