Skip to content

Arabic AI and document assistants

Arabic AI that answers from your documents, with the source.

ARO Solutions builds Arabic document assistants: retrieval-augmented generation (RAG) systems that answer questions from Arabic contracts, policies, regulations and records, including scanned files, and show the exact source passage for every answer. The studio is based in Amman, Jordan, works natively in Arabic and English, and offers a first assistant as a fixed-price pilot of about three weeks, to companies directly or white-label to IT integrators.

Last updated · ARO Solutions, Amman

Why Arabic documents break ordinary RAG

Most retrieval pipelines are designed and tested on clean English text. Arabic archives are different in ways that decide whether the assistant can be trusted:

  • Scanned pages

    Much of the Arabic record in banks, insurers and law firms is scanned. Arabic OCR drops dots, joins or splits letters and scrambles tables and mixed Arabic-English lines, so the text layer is checked and corrected before anything is indexed.

  • One word, many spellings

    The same word appears with أ‎, إ‎, آ‎ or ا‎, with ة‎ or ه‎, with ى‎ or ي‎, with or without diacritics and elongation. Without normalisation, search misses documents a person would find at once.

  • Words that carry their grammar

    Arabic attaches articles, prepositions and pronouns to the word itself (والعقد، بعقودهم). Keyword search needs Arabic-aware analysis working alongside multilingual embeddings, not one or the other.

  • Dialect questions, formal documents

    Staff and customers ask in Gulf or Levantine dialect; the documents are in Modern Standard Arabic. The question is rewritten into the register of the documents before retrieval.

  • Mixed languages, digits and dates

    Contracts mix Arabic and English, Arabic-Indic digits (٠١٢٣) and Western ones, Hijri and Gregorian dates. Each is normalised so a clause number or a date matches however it was typed.

  • Articles and clauses

    Legal and policy texts are organised by article (المادة), clause and annex. Splitting follows that structure, so an answer cites Article 14 on page 9 rather than an arbitrary slice of text.

What you get from the pilot

  • An assistant that answers from an agreed set of your documents, in Arabic and English.

  • The source passage and page shown with every answer, in a right-to-left interface.

  • A plain "the documents don't say" when the answer isn't there, instead of a guess.

  • An evaluation report: accuracy on a test set of questions written with your team, for example 100 of them.

  • Deployment in your environment or your client's, with the code in your repositories.

  • Documentation and a handover session, plus a costed plan for the next phase.

How the three weeks run

  1. Week one: documents in, the OCR and cleaning pipeline, a first index, and the test questions agreed with your team.

  2. Week two: retrieval tuned on your documents, answers generated with citations, and a first demo to real users.

  3. Week three: the evaluation run against the test set, fixes, access rules and handover.

Typical uses

  • Banks and insurers

    Policy wording, product terms, compliance circulars and internal procedures, answered for staff in seconds.

  • Law firms and legal teams

    Contracts, case files and legislation, searched by meaning and cited by article.

  • Real estate and construction

    Tenancy contracts, tender documents and specifications, checked and queried without reading every page.

  • HR and operations

    Policies, procedures and handbooks, so employees ask the assistant before they ask a colleague.

  • Customer service

    Answers from your knowledge base in Arabic on web chat, with a hand-off to a person when it matters.

Three ways to get an Arabic assistant

CriterionOff-the-shelf chatbotRAG built for English firstArabic-first RAG (what ARO builds)
Scanned Arabic PDFsOften not supportedOCR added late; its errors pass into the indexOCR output checked and corrected before indexing
Dialect questions on formal documentsHit and missOften missesQuestion rewritten into the documents' register
CitationsRareTo a chunk of textTo the article, clause and page
Where the data livesThe vendor's cloudDepends on the buildYour environment or your client's
Proof it worksA vendor demoSpot checksAn accuracy report on your own test questions

Swipe the table sideways to see every column.

What we build with

  • RAG
  • Arabic OCR
  • Multilingual embeddings
  • Hybrid search
  • Claude API
  • LangGraph
  • Postgres
  • Python

Questions people ask

What is an Arabic document assistant?

It is an AI assistant that answers questions from your own Arabic documents (contracts, policies, regulations, records) rather than from general knowledge, and shows the passage it relied on. Technically it is retrieval-augmented generation (RAG): the system finds the relevant passages first, then a language model writes the answer from them.

Can it read scanned Arabic PDFs and handwriting?

Scanned printed Arabic, yes: the pipeline runs OCR, then checks and corrects the output before indexing, because raw Arabic OCR is where most answer errors start. Handwritten Arabic is much less reliable. We test it on your own samples and tell you the result before you commit to anything.

Does it understand Gulf, Levantine and Egyptian dialects?

Yes for questions. People can ask in their own dialect; the system rewrites the question into Modern Standard Arabic, which is how the documents are written, before it searches. Answers come back in clear formal Arabic or in English.

Which models do you use for Arabic?

Claude and other LLM APIs for writing the answer, and multilingual embedding models for retrieval, chosen by testing them on your documents rather than by a leaderboard. When data must stay in the country, we use in-region or self-hosted models.

Where is our data stored? Can it stay in the UAE or Saudi Arabia?

The documents and the index stay in your environment or your client's. We document which LLM APIs are called and their data-retention settings, and we can use in-region or self-hosted models when a client's rules require data to stay in the country.

How accurate is it?

We don't quote a generic number, because accuracy depends on your documents. It is measured on a test set of questions written with your team, and the report is part of the handover. When the documents don't contain an answer, the assistant says so instead of guessing.

How long does it take, and what does it cost?

A first assistant on an agreed set of documents is a fixed-price pilot of about three weeks. The price is confirmed in writing within 48 hours of a scoping call. IT integrators and agencies can ask for the partner rate card.

Can you deliver this under an integrator's brand?

Yes. ARO takes this work white-label from IT integrators and vendor partners: you win and keep the client, we build the Arabic AI part under your name with non-circumvention in writing, and you get a free solution outline for one live bid within 48 hours.

Got an idea?
Let's build.

Tell us what you want to build. We usually reply within a few hours. Worst case, by the end of the day.

Amman, Jordan · Taking projects worldwide

Free intro call

Not sure what you need yet? Book 30 minutes — we'll figure out what's actually worth building.