AI knowledge base
(RAG)
AI knowledge base (RAG)
An assistant that answers questions about your company's documentation by citing the source, within each user's permissions and without straying beyond what is written.
An AI knowledge base connects to wherever your documents already live: SharePoint, Drive, the document management system, network drives. It lets people ask them questions in plain language. The notice period in a supplier contract. The procedure when a quality test fails. Whether that incident had happened before and how it was resolved. The answer arrives in seconds, with the passage it comes from and a link to open the document.
The technique behind this kind of document AI is called retrieval augmented generation, or RAG. Before answering, the system retrieves the relevant passages from your documentation and forces the model to answer only from them. That is why it is not a generic chatbot: ChatGPT does not know your contracts; this assistant knows nothing else. And when the answer is not in the documents, it says so.
Three questions that always come up, answered plainly.
It answers from your documents, not from the internet, and shows you where every sentence comes from. It also controls which documents each person can see, and what it does when there is no answer: say so.
Yes, if there is documentation that many people consult and only two or three really know: contracts, procedures, technical data sheets, regulations. The more the company depends on those people, the sooner it pays for itself.
The first useful assistant, on a defined set of documents, takes 4 to 8 weeks. We start with the folder that generates the most questions, not by indexing everything.
Six pieces that work together. The last three are what separate an impressive pilot from a system you can rely on.
Your team asks in their own language and gets the answer with the exact passage and a link to the document. It sits inside Teams, Slack or your intranet.
SharePoint, Google Drive, the document management system, email, network drives and the ERP if needed. The documents do not move: the assistant reads them where they live.
Each person only gets answers from the documents they could already open. Permissions are inherited from your directory, and there is a record of who asked what.
A dashboard of unanswered questions, badly rated answers and missing documents. It is what makes the team's trust grow instead of breaking.
When a document changes, the old version stops answering. Scheduled synchronisation, duplicate detection and alerts for documents that are out of date or have no owner.
In your cloud (Azure, AWS, Google) or on your own servers, with closed or open models depending on how sensitive the documents are. You decide.
Five steps, and the first is deciding which documents do NOT go in yet.
We choose a defined set with you: supplier contracts, the quality manual or the incident history. Not the whole server.
Formats, versions, duplicates, scans with no text layer and current permissions. This gives us the real timeline and tells us what needs cleaning before indexing.
Within weeks, using the questions the team actually asks. We measure how many it answers well, how many it cannot answer and why.
We improve retrieval, add the missing documents and fix the cases that fail. With the team in the room, not back at our office.
Documented, and with your team able to add sources. Then, if you want, more departments, or the step up to agents that act on what the assistant knows.
Paste your procedure into ChatGPT and ask it a question. The answer comes from three places at once and does not tell you which: partly from what you pasted, partly from what it read on the internet when it was trained, and partly from what it made up so the sentence would fit. All three arrive in the same confident tone.
An AI knowledge base cannot do that. Before answering, it looks for the passages that deal with your question inside your documents and forces the model to answer only from them. That is the technique the industry calls RAG. It is why every answer comes with the document, page and version it is drawn from, and why, when the answer is not in your documents, it says so instead of filling the gap.
The practical difference shows in what you can do next. A ChatGPT answer has to be checked before you use it, and checking it costs as much as finding it yourself would have. An answer with its citation can be checked in ten seconds by opening the document, and that is the only thing that turns this into a working tool.
| ChatGPT | AI knowledge base (RAG) |
|---|---|
| Three sources mixed up | Only your documents |
| No citation | Document, page and version |
| Fills in the gaps | Says it isn’t there |
Procedures, work instructions, ISO standards and technical data sheets. Someone on the shop floor asks and gets the exact step from the current version, without searching the server or phoning the office.
Clauses, deadlines, renewals and penalties in customer and supplier contracts. This is where the automatic renewals nobody had noted down come to light.
Product manuals and the history of resolved incidents. A technician in their first month answers like one with ten years behind them, because both consult the same resolutions.
Previous proposals, tender specifications, certificates and references. Preparing a proposal is no longer a hunt through the folder of the last similar one.
Collective agreements, internal policies, protocols and staff announcements. The questions HR hears again and again answer themselves, with the source, and only for those allowed to see it.
The doubts that appear as soon as anyone mentions touching the company's documents. If yours is missing, write to us.
Let's talk about your projectOnly if you decide they should. It can be deployed in your cloud, in ours or on your own servers. With models that are contractually barred from training on your data, or with open models inside your own infrastructure.
It only answers from passages in your documents, and it shows them. If it cannot find the information, it says so. We measure the accuracy rate with real questions before opening it up to the team.
They are synchronised at the agreed frequency, usually every night. The new version replaces the old one; if a document is deleted, it disappears from the assistant.
There are two parts. The project depends on the state of the documentation and on how many sources need connecting. Usage is paid per query, and it is estimated from your real volume before anything is signed.
Digital documentation, and someone who knows which version is the right one. We process the scanned documents; the audit brings the duplicate versions to light.
The assistant answers; the agent acts. An agent that handles incidents needs to know how the previous ones were resolved, and the assistant is what provides that. It is the first step, and the safest one.
It is the entry point in construction and engineering: a tender document can run to three hundred pages, with the conditions that decide the margin scattered through it. It comes second in yachting and marine, with haul-out reports, job sheets and refit specifications. And it also leads in professional services, the only industry where the document does not come with the product: it is the product being sold.
