AI & Automation
RAG AI assistants in Qatar
An assistant that answers from your documents, with the source attached.
Warranty termsv3
1. CoverFacade stone is covered against cracking and colour loss for 10 years from handover.
2. ExclusionsDamage from building settlement, impact or unapproved cleaning products is not covered.
3. ClaimsSend photos and the project number to the office. A site visit follows within 5 working days.
10 years from handover, against cracking and colour loss.
From: 1. Cover
7
projects shipped across Qatar and the GCC
Design, search visibility and pipeline — built together, in English and Arabic.
Definition
What retrieval-augmented generation means
Retrieval-augmented generation, RAG for short, makes an AI model answer from your documents instead of from memory. The assistant first searches your own files, such as policies, specifications, price lists and contracts, pulls out the passages that match, and writes an answer from those alone, citing them. If nothing matches, it says so.
It is not a model trained on your data: change a document and the next answer changes with it. And unlike a chatbot, which covers a short list of public questions, it searches hundreds or thousands of pages, often internal ones.
Why it matters
Why grounding is the whole point
A model with no source gives a confident, plausible answer whether or not it knows, which is unusable for anything that matters. Retrieval ties every answer to a document you control, so a person can check the citation in seconds and answers stay current when the source changes. Try it above: move the warranty terms to version 4 and the answer follows.
Documents in two languages
An Arabic contract, an English specification, a policy in both. The question can be asked in either and still find the passage.
Knowledge held by one person
Answers stop waiting for the senior colleague who is on leave or on site, and new staff stop asking where things are.
A source you can check
Every answer points to its passage, so the person who relies on it spends seconds checking, not minutes searching.
A person still signs
Anything that leaves the business, a reply, a quote or a contract position, is approved by a person first.
Under the hood
How an answer is put together, and where it goes wrong
Every RAG assistant runs these five steps, and most wrong answers start in the first two, not in the model.
Split
Documents are cut into passages. Cut badly, a clause loses the sentence that qualifies it. Ask how tables and numbered clauses are handled.
Index
Each passage is stored by meaning, so it is found across wording and language. Scanned PDFs with no text must be converted first.
Retrieve
The closest passages are pulled, filtered by what the asker may see. Permissions belong here, before the model, not in its instructions.
Answer
The model writes only from those passages and cites them. If none is close enough, the answer is: not in the documents.
Check
A fixed set of test questions with known answers runs after every change, so a broken update shows at once.
The limits
What the assistant does, and what stays with a person
RAG makes errors rarer and checkable, not impossible, so the line between answering and sending is drawn in advance.
The assistant does
- Answers staff, with the sourceA colleague reads the cited passage before relying on it, a matter of seconds.
- Drafts repliesCustomer replies, quotes and contract positions are written as drafts, never sent.
- Says when it doesn't knowNo close passage means no answer, and a pointer to who can help.
A person does
- Approves anything that leavesNothing with your name on it goes out until someone has read it.
- Decides what never goes inPayroll, staff files, patient records, ID copies and client confidences. What cannot be retrieved cannot leak.
- Chooses the providerRead its terms on training, retention and region. Under Law No. 13 of 2016, personal data needs a lawful basis, usually consent.
An example
An example: warranty questions at a stone-cladding firm
An example, the same firm as the demo above, not a client.
What goes in
Current warranty terms, installation specifications and maintenance guides, in Arabic and English. Old versions removed; quotes and customer files left out.
A supervisor asks
In Arabic, about the spacing of facade fixings. The answer comes from the English specification and quotes the line.
Customers ask
A crack from the building settling? Not covered, citing the exclusions. A discount? In no document, so the office will answer.
The terms change
Cover moves from 10 to 12 years. The office uploads version 4, and the test question confirms the new answer.
Replies go out
Answers to customers stay drafts until someone in the office sends them.
Before you start
Getting your documents ready
Worth doing whether or not you build an assistant; a tidy shared drive often solves half the problem.
- One current version of each documentArchive the rest where the assistant cannot read them.
- Text, not pictures of textIf you cannot select a sentence in a PDF, it is a scan and needs converting.
- Write down what one person knowsA short written note is the document the assistant needs.
- Decide who sees whatEveryone, staff only, managers only, nobody. That list becomes the access rules.
- 20 real questions with known answersYour test before launch and after every update.
- An owner for the documentsWithout one, answers stay right only until the first price change.
What RAG assistants includes with us
A document inventory: what goes in, what stays out, and who may see each
Files taken from where they live now: shared drives, PDFs, your website, your systems
A citation on every answer, so anyone can check the passage in seconds
"Not in the documents" when nothing matches, instead of a guess
Questions in Arabic or English, answered from documents in either language
Access by role, applied before the model sees anything
Test questions with known answers, re-run whenever the documents change
A refresh routine that replaces old versions instead of keeping both
Common questions
How much does a RAG assistant cost?
It is quoted after a short call, once we have seen the documents. What sets the scope: how many there are and what state they are in, whether scans need converting, the languages, who uses it and with what access, and which systems it reads. Running costs sit on your own accounts: the AI provider bills per question, and storage and hosting are on your plan.
Where does our data go?
We agree that in writing before building: which AI provider processes the questions and passages, in which region, what it retains and for how long, and whether it may train on the data. Terms differ by plan and change over time, so read the current ones. The documents stay in storage on your account, and anything you would not send to an outside provider stays out of the index.
Can it still give a wrong answer?
Yes. It can misread a passage, merge two clauses, or cite the right document and summarise it badly. Grounding makes errors rarer and checkable, because every answer shows its source. The test questions catch regressions after each update. That is why anything going to a customer or a counterparty is a draft a person approves, never a message the assistant sends.
How current are the answers?
As current as the documents. Replace a file and the next answer uses the new version, as the demo above shows. The real risk is duplication: if the old version stays where the assistant can read it, both are retrieved and the answer may mix them. So the refresh routine removes the old file and re-runs the test questions.
Can customers use it, or only staff?
Both, as separate assistants with separate documents. The customer one reads only what you would publish, such as warranty terms and care guides. The staff one also reads internal procedures, filtered by role. Keeping them apart is safer than telling one assistant to hide things, because instructions can be talked around, while a document not in the index cannot be retrieved.
Do we need RAG, or would a chatbot do?
If customers ask a short list of questions your website already answers, a scoped chatbot is simpler. RAG earns its place when answers live in many long documents, change often, or need a source someone can check. A quick test: if the person who answers today opens a file to find the answer, RAG fits. If they answer from memory, a chatbot probably does.
What do you need from us?
Access to the documents, an hour with the people who answer the questions today, a list of real questions with known answers, and a decision on who may see what. After launch, one person owns the documents and replaces them when they change. If you have no AI provider account yet, we open one in your company's name.
More in AI & Automation
Every part of AI & Automation works better alongside the others — here's the rest of it.
Ready to hand the busywork over?
15 minutes · A clear next step · No pressure