Collect the Real Questions
Questions from tickets, inboxes and chat become the test set before code is written.
Answers with the source attached.
AI search over your own documents, for your staff. AI that answers questions from your own documents, tickets, records, policies and knowledge base. This is known as RAG (retrieval-augmented generation). It works inside the tools your staff already use, links its sources and respects who is asking.
Documentation only a few pages long? A clear help centre is cheaper and solves most of it.
Honest before clever
Next in Build
WordPress & WooCommerce
Stores built well, kept running, with AI inside.
What it is
Answers from your own files, with the source attached
Retrieval over private data (RAG) is AI that answers your staff's questions from your own documents, tickets, records and policies, inside the tools they already use.
Every answer links to its sources and respects what the person asking may see. When your data has no answer, it says so plainly. We use the same approach in our own product: the Real Estate Club Dubai advisor answers by looking up its own database and returns source links.
Retrieval Over Private Data answers your staff from internal documents behind a login. Customer Assistants answer customers. Agent Pipelines and Process Automation & Integrations take actions in your systems. Letting AI apps such as Claude reach your live systems directly, through an MCP server (Model Context Protocol, a standard way to give AI tools controlled access), is part of our AI Systems work.
If the content is small and stable, a well-organised help centre with plain search does most of the job for far less. Training a model on your documents (fine-tuning) is rarely the fix either. It cannot show sources or forget a withdrawn policy.
What walks in
Four signs your answers are stuck in documents
The answers exist somewhere in your files. The problem is finding them, trusting them and keeping them once the people who know leave.
New staff ask the same questions every week, and the shared drive cannot answer them.
Start With a Free Meeting
02Two people know where everything is. One of them is leaving.
Plan What Comes First
03A chat-with-your-documents tool answered from things that were not there.
Get a System Review
04Customers, not staff, ask the same questions every day.
Customer Assistants
What we already know
The details that decide whether it holds up
-
01
Plan for itMost failures are search failures, not AI failures
If the right passage is not in the top results, no model can answer correctly. So we measure search before building the answer step.
-
02
Permissions must filter before the model sees anything
Filtering the answer afterwards is too late. The search itself is filtered by the roles the person already has in your system.
-
03
Deleted and replaced documents are the classic leak
An index that only ever adds keeps answering from last year's price list. So deletions and ‘current version’ rules are part of the sync.
-
04
Documents can carry instructions
A file can contain text written to trick the model (‘ignore your rules and…’). So text found in documents is treated as data only and cannot trigger actions.
Who it is for
Worth building when the answers are written down and permissions matter
Training a model on your documents (fine-tuning) is rarely the fix either. It cannot show sources or forget a withdrawn policy.
A good fit
-
01
The knowledge your team needs is written down somewhere: PDFs, policies, manuals, tickets, product data, a wiki.
-
02
Different people may see different documents, and any search you add has to respect that.
-
03
Staff waste time searching or asking the same senior people, and wrong answers have a real cost.
Not a fit yet
-
Documentation a Few Pages Long
A well-organised help centre with plain search does most of the job for far less.
Organise it into a help centre or wiki first
-
Answers Not Written Down, or Wrong
The AI repeats what it finds, mistakes included.
Fix the content first
How a job starts here
Real questions first, then sources, then answers
Where sources and permissions need mapping, it starts with a System Review, delivered within a week. Then one fixed price before any code is written.
Map Sources and Permissions
We list every source and who may see what. We agree in writing what data may go to the AI provider.
Index One Source, Measure, Then Answer
Most failures are search failures. So we tune how documents are split and searched against the test set, before adding answers.
Ship into One Tool, Then Run It
On a Run plan we close unanswered gaps, add sources and re-run the test set on every change.
In production
We run the same pattern in our own products
FAQ
Questions people ask before searching their own documents
Yes: this is RAG (retrieval-augmented generation) over your own knowledge base. Staff ask in plain language and get answers from your documents, with sources. Your documents and the index stay on your own servers. Only the passages a question needs go to the AI provider you approved.
It answers only from the passages it found, shows them as sources, and says ‘I don’t know’ when nothing relevant turns up. That cuts down invented answers. Where a wrong answer is costly, staff should still open the source.
Not if your source permissions are right. If they are wrong today, the system repeats the mistake, which is why we map them first.
Usually a fixed-price build. The price depends on how many sources there are, how messy they are, and the permissions. Running it is a monthly Run plan, and model usage is separate, under a spending limit. The current figures are on our pricing page.
