Kashfara is a compact 4-billion-parameter multimodal model purpose-trained for document and OCR understanding — combining fast parsing with vision reasoning for low-compute, high-accuracy Arabic & English automation.
Instead of brute-forcing raw pixels with heavy compute, Kashfara parses structure first and pairs it with targeted multimodal vision reasoning.
PDFs and Office files are parsed for native text layers and page geometry in milliseconds, keeping token usage lean and responses fast.
Preserves Arabic RTL script, English LTR text, tables, checkboxes, stamps, and handwritten signatures across structured and unstructured layouts.
Every extracted field links back to its source page, verbatim quote, and normalized bounding box — zero blind guesswork.
Automate Arabic-English tax invoices, bank statements, purchase orders, and vendor receipts with line-item table extraction.
Extract audited balance sheets, trade finance documents, commercial registrations, and corporate onboarding forms in seconds.
Extract clauses, parties, dates, stamps, and handwritten signatures from mixed RTL/LTR agreements and HR forms.
Turn multi-page shipping manifests, delivery notes, and packing lists into structured ERP-ready JSON and CSV exports.
Process bilingual tenancy contracts, maintenance reports, and property inspection checklists with proof-of-evidence coordinates.
Digitize insurance claims, policy schedules, prescriptions, and patient intake forms with high accuracy across handwritten and printed text.
Upload any Arabic, English, or bilingual PDF, image, or Office document in the Live Demo to inspect extracted fields, tables, digital forms, and grounded Q&A.