RAG Knowledge System.

Ask a general AI about your business and it will give a plausible answer. Plausible is not good enough when the question is about your prices, your policies, or what you actually did on a past job. A RAG knowledge system changes the order: look up your material first, answer from that, and say so when the material does not cover the question.

What it actually covers

Collecting and cleaning the source material: website, service sheets, price lists, policies, FAQs, past project notes. Building the retrieval layer so a question pulls the right documents first. An assistant on top — staff-facing, customer-facing, or both — with clear limits on what it can promise. No outcome promises, and no regulated claims without a human check.

The review loop is part of the system: wrong answers get corrected at the source document, not patched over with a one-off reply. If the source is wrong, every future answer will be wrong until the source is fixed.

• Source material collection and cleanup
• Retrieval layer over your own documents
• Staff-facing and/or customer-facing assistant
• Limits on promises and regulated claims
• Correction loop at the source document

Where it helps most

New staff who need reliable answers without interrupting a senior. Intake teams who field the same twenty questions. Quote preparation where the right past job or price list has to be found quickly. And customer-facing Q&A where a fast, accurate answer is the difference between an enquiry and a bounce.

Isolation, by design

Each client's information and context are containerized and isolated, never blended or reused across clients. A knowledge system that mixes one business's material into another's answers is not a feature — it is a breach of trust, and it is designed out from the start.

What people are actually looking for

Staff and customers ask the same questions over and over: what do you offer, what does it cost, what is your policy, what did you do on a job like this, how do I do this task? The answers exist — in a website, a price list, a policy document, a past project file — but finding them takes a senior person, an interruption, and time.

A RAG knowledge system is the work of making those answers findable, from your own material, in seconds, without guessing.

What is working in 2026

Retrieval before generation. The AI looks up your documents first, then answers from them. That single change is the difference between a plausible answer and a correct one, and it is why RAG — retrieval-augmented generation — has become the standard pattern for business AI assistants, instead of asking a general model to guess about your business.

Staff-facing first, customer-facing second. The fastest return is usually internal: new staff, intake teams and quote preparation, where a reliable answer saves an interruption. Customer-facing Q&A comes next, once the internal system is proven and its limits are known.

Correction at the source. When the system gives a wrong answer, the fix is to correct the source document, not to patch the reply. If the source is wrong, every future answer will be wrong until the source is fixed. That loop is part of the system, not an afterthought.

Isolation per client. Each client's information and context are containerized and isolated, never blended or reused across clients. A knowledge system that mixes one business's material into another's answers is not a feature — it is a breach of trust.

What we have done

We have built knowledge systems around a business's own material — website, service sheets, price lists, policies, FAQs and past project notes — with a retrieval layer that pulls the right documents first, and an assistant on top for staff and, where appropriate, customers.

The use cases that have proven most useful are the unglamorous ones: intake support, where the same twenty questions need reliable answers; quote preparation, where the right past job or price list has to be found quickly; and training, where new staff can ask without interrupting a senior every time.

What you get

Source material collection and cleanup. A retrieval layer over your own documents. An assistant — staff-facing, customer-facing, or both — with clear limits on what it can promise. A review loop for wrong answers, corrected at the source. And documentation that stays with the business.

• Source material collection and cleanup
• Retrieval layer over your own documents
• Staff-facing and/or customer-facing assistant
• Limits on promises and regulated claims
• Correction loop at the source document
• Per-client isolation, by design

How AI is used — and where it is not

AI is the whole point here, but it is constrained AI: it retrieves your documents, summarizes them, and answers from them, with sources the user or staff member can check. It does not answer from general knowledge when your material covers the question, and it says so when your material does not cover it.

It is not used to invent a price, a policy, a past result or a regulated claim. Where a question needs a human — a clinical judgement, a complex quote, a complaint — the system hands it over, with the retrieved material attached, instead of guessing. And no client's material is ever used to answer another client's questions.

Questions people actually ask

What is RAG, in plain terms?

Retrieval-augmented generation. Before the AI answers, it retrieves — looks up — your own documents that are relevant to the question, and then generates its answer from those documents. Without retrieval, a general AI answers from what it was trained on, which may be plausible and wrong about your business. With retrieval, it answers from your material.

What material do I need to start?

Whatever you already have: website, service sheets, price lists, policies, FAQs, past project notes. The first step is collecting and cleaning that material — removing duplicates, fixing outdated pages, and filling the gaps where a common question has no document answering it.

Can customers use it, or just staff?

Both, in that order. Staff-facing first is usually the fastest, safest start: it proves the system, surfaces the gaps, and builds the correction loop. Customer-facing Q&A comes next, once the limits are known and the handover to a human is clean.

Will it hallucinate?

Less, if it is built right — and never silently. Retrieval, source citations, and a clear 'your material does not cover this' response are the controls. But no AI system should be left unsupervised on regulated or high-stakes questions. Human checks stay where the answer matters.

Is my data shared with other clients?

No. Each client's information and context are containerized and isolated, never blended or reused across clients. Your documents answer your questions, and nobody else's.

Do I own the knowledge base?

Yes. The knowledge base, the retrieval setup and the documentation are yours, on infrastructure you control.