How to prepare your documents for an AI knowledge base: a checklist for UAE companies

TL;DR: An AI knowledge base answers from the documents you give it, so preparing those documents is most of the project. Inventory what you have, archive what is stale, keep one current version of each document with a named owner and a date, make scanned files readable, keep prices and rules in structured tables, pair Arabic and English where customers see it, remove personal data that does not belong, and write a set of test questions with the right answers. None of this needs an engineer. All of it decides whether the assistant is trustworthy.

Why the documents matter more than the model

Preparing documents for an AI knowledge base means turning the pile of files your company runs on into a set the AI can read and trust: current, deduplicated, owned, machine-readable, and free of anything that should not be there. A retrieval-based assistant does not know your business; it looks up your documents and answers from what it finds. If the documents contradict each other, the answers will too.

This is the step most UAE companies underestimate. The model choice is a short conversation. The documents are weeks of quiet editorial work, most of which can only be done by people inside the business, because only they know which of the four versions of the leave policy is the real one. This checklist is meant to be worked through by an operations manager or office lead before an engineer is involved, and it applies whether the assistant is for customers on WhatsApp or for staff searching internal manuals.

Step 1: Inventory what you have

Start by listing the documents that hold answers, not every file on the server. The practical test is: when a customer or a new employee asks a question, what do experienced staff open or remember? That set is your first knowledge base.

• Service and product descriptions, specifications and catalogues
• Price lists, rate cards and packages
• Policies customers see: cancellation, refunds, warranty, delivery, terms and conditions
• Internal procedures: how to book, how to onboard, how to escalate
• Frequently asked questions, formal or informal, including the replies staff keep pasting into WhatsApp
• Contracts and templates, if the assistant is meant to answer about them
• Manuals, safety documents and compliance procedures

For each document, note where it lives (shared drive, email, someone's laptop, a printed binder), the format, and who last changed it. The output is a simple table, and the gaps in it are the first finding: usually a few critical answers exist only in one person's head. Those need to be written down before anything else.

Step 2: Decide what is current

Every company has a folder called "Final", a folder called "Final v2", and a document called "Final FINAL". The AI cannot tell which one is real. Before loading anything, resolve every duplicate to a single current version, archive the rest somewhere the assistant does not read, and make sure the current version matches what the business actually does today, not what it did when the document was written.

Then give each document two pieces of metadata: an owner (a named person, not a department) and an effective date. The owner is the person who will be asked when the assistant says something wrong, and who reloads the document when the fact changes. The date lets you see at a glance what has not been reviewed in a while. This sounds bureaucratic; it is the single habit that keeps a knowledge base accurate after launch.

Step 3: Make everything machine-readable

An AI knowledge base searches text. A scanned PDF, a photograph of a price list, or a screenshot of a policy is an image, and images are not searchable until they are converted. The conversion is called OCR, and it works, with caveats an engineer can defend: it is less accurate on poor scans, weaker on Arabic than on English, and often garbles tables. Where the original Word or Excel file exists, use that instead of the scan. Where only the scan exists, run OCR and have someone check the result, especially numbers.

Format matters beyond scans. Documents with clear headings are split into sensible pieces for retrieval; a single wall of text is not. If a long document covers many topics, add headings that say what each section is about. Avoid answers that exist only in a diagram, a footer, or a colour code, because the assistant will not see them.

Step 4: Put facts in tables and explanations in prose

The most common source of confident wrong answers is a fact buried in a paragraph. A price mentioned in a sentence can be paired with the wrong item; a delivery rule described across three paragraphs can be half-read. Anything that must be exact should live in a structured table with one row per item and clear columns: item, code, unit, price, effective date, notes. The same applies to opening hours by branch, service areas, eligibility rules, and delivery zones.

In a well-built system, those tables are looked up rather than interpreted: the assistant fetches the row rather than reading around it. That only works if the table exists. Prose is still right for explanations, policies and procedures, where the reader needs context rather than a number.

Step 5: Handle Arabic and English deliberately

UAE customers write in both languages, often in the same message. A modern model can usually answer an Arabic question from an English document, and vice versa, but three things go wrong without care. Names, product terms and place names get translated when they should be kept. Legal and policy wording shifts meaning in translation. And an assistant that answers in fluent Modern Standard Arabic when the customer wrote in Gulf dialect reads as stiff, even when it is correct.

The practical rule: for documents customers see, especially policies, terms and service descriptions, have a person write the Arabic version and store it paired with the English one, so the assistant can retrieve the same fact in either language. For internal manuals, one language is usually fine. Where terminology matters, keep a short glossary of terms that must not be translated. We go deeper into the engineering side in our piece on why Arabic breaks language models.

Step 6: Remove what should not be there

Documents accumulate personal data: customer names and numbers in email threads, Emirates ID scans attached to contracts, salaries in HR policies, patient details in clinic procedures. An assistant that answers customers must not have access to any of it, and an internal assistant should see only what each role is allowed to see.

Before loading, redact or remove personal data that the assistant does not need to do its job. Keep sensitive material, such as contracts with named parties or medical records, in a separately controlled store with role-based access, or leave it out entirely. Check the current UAE data protection guidance for your sector rather than assuming; healthcare and finance carry additional rules. Where the data must stay in the country, the hosting decision belongs in the same conversation, which we cover in AI data residency under the UAE PDPL.

Step 7: Write the test questions

Before the assistant exists, write down the questions it will be asked, with the answers you expect and the document each answer comes from. Ask the people who answer the phone and the WhatsApp number; they know the real questions, including the awkward ones. Aim for a mix: routine questions, questions with exact answers such as prices, questions that should be handed to a person, and questions in both languages.

This list has two uses. It tells you which documents are missing before the build starts, because a question with no source document is a gap. And it becomes the acceptance test: the assistant is ready when it answers the list correctly and hands off the ones it should. Without it, "does it work?" is a matter of opinion.

The checklist

Item	Done when	Who
Inventory of answer-bearing documents	A table listing each document, where it lives, its format and last editor	Operations lead
Knowledge that lives in heads	Written down as documents with the same metadata	Senior staff
One current version per document	Duplicates and old versions archived out of the assistant's reach	Document owners
Owner and effective date on each document	Every document has a named person and a date	Operations lead
Scans converted to text and checked	No image-only documents remain; numbers verified after OCR	Admin, with engineer support
Headings on long documents	Each section says what it covers	Document owners
Facts in tables	Prices, hours, zones and rules in structured rows with dates	Finance or operations
Arabic and English pairs for customer-facing documents	Human-written Arabic stored alongside English; glossary of untranslatable terms	Bilingual staff
Personal data removed or separated	No customer, employee or patient data in the general knowledge base	Owner, with sector guidance checked
Access rules for internal documents	Each document marked with who may see it	Management
Test questions with expected answers	A list covering routine, exact, handoff and bilingual questions, each with its source	Front-line staff
Update process	Each owner knows how to reload their document when a fact changes	Operations lead

How Soluvide approaches this

We start every knowledge base project with this checklist, because the build is fast once the documents are right and slow forever if they are not. The engineering side, retrieval over your documents, table lookups for exact facts, per-role access, and hosting where the data has to stay, is what we deliver under private knowledge bases and RAG. When the same knowledge needs to reach customers on WhatsApp or the website, it becomes the grounding for an assistant built under AI chatbots and agents.

If you are not sure which documents matter or where the gaps are, our free automation audit is the right first step: we walk the inventory with you and tell you what is ready and what is not. Scope first, then a fixed-fee proposal.

The fastest way to find out how much preparation you need is to send us three things: your price list, your most-asked customer question, and the document you would use to answer it. Message us on WhatsApp and we will tell you what we see.