00 AI knowledge base
Somewhere in your policies, manuals and shared folders is the answer your team keeps being asked for. We build the assistant that finds it, answers in plain words and shows the passage it came from. Built in the UK for firms where a wrong answer costs more than a slow one.
01 Where it helps
Staff or clients ask what a policy already says. The one person who knows ends up giving the answer by hand, again.
A 600-page handbook, three years of meeting notes, a folder of PDFs. Nobody can ask it a question the way they would ask a colleague.
When the policy lives in someone's head, two colleagues give two answers. An assistant that answers from the document gives one.
A public page that answers routine questions from approved text, and hands anything personal or unclear to your team with the conversation attached.
Compliance, legal, healthcare, finance, lettings. A fluent wrong answer is worse than none, so this one shows its source or says it does not know.
Prices, contracts, regulation. Update the document and the answers follow, within minutes or overnight depending on the source. Nothing to retrain.
02 What it is
A general-purpose AI assistant does not know your prices, your contracts, your internal policies or last week's board minutes. Ask it about those and you get a fluent guess. That is the wrong way to fail for a firm answering client questions, a clinic checking a procedure or a lettings team explaining a policy to landlords.
An AI knowledge base works differently. When a question arrives, it searches a private library of your documents first, pulls out the passages that matter and only then writes the answer from them. Every answer can show the paragraph, file or policy version it came from.
It is not trained on your documents. It looks them up every time, so there is nothing to untangle when a document changes or has to be removed. We build the plumbing, so the person asking sees a plain answer and a source link: here is the answer, here is where it came from, and here is where your documents stop.
03 When it does not know
01 · Covered
When your approved text covers the question, you get a short answer in your own tone and the passage it came from, so anyone can check it.
02 · Nearly
When the answer depends on a fact it does not have, it asks for that fact instead of guessing.
03 · Not covered
When your documents do not cover it, it says that plainly and passes the person to a human with the conversation attached. A refusal becomes a better handover, not a dead end.
One client in a regulated services sector put it this way: "The useful part is not that it answers everything. It is that it knows when to stop, and our team can see exactly what happened." (UK, anonymised.)
04 When not to buy this
Uploading a pack of documents into a chat assistant is fine for a one-off job, such as reading one long contract. It is much less useful when a business needs the same dependable answers, day after day, over thousands of changing files, with different people allowed to see different things.
Plain search returns documents. A knowledge base returns an answer, says what it was based on and links back to the evidence. Retraining a model can teach it a tone or a format, but it is a poor way to keep a policy library current. If the handbook changes on Friday, we would rather update the document than retrain behaviour into a model and hope everyone remembers which version is live.
And if what really eats your week is repetitive admin rather than finding answers, that is AI automation. If you are not sure which of the two you have, our AI consultancy starts with how the work actually runs.
If a knowledge base is not the answer, because you want a chatbot that merely sounds friendly, your data fits in a single prompt or you need a change of tone, we will tell you before we build the wrong thing. Here are three kinds of AI projects we turn down.
05 How we build
01 · Source audit
Shared drives, wikis, databases, a folder of PDFs. We read every format and flag the documents that need cleaning before they are useful.
02 · Search design
We choose how your documents are split, tagged and searched to fit the material, not the other way round. A long policy reads differently from a flat product list.
03 · Answer rules
Whether it answers, asks a follow-up or refuses and hands over to a person, with a source on every answer. Most failed knowledge base projects break here.
04 · Operations
A web widget, a team chat bot, a WhatsApp number or an internal admin tool. Then logging, a test set, an alert when answers get worse and a refresh schedule that matches your sources.
06 What we check
We collect real questions from staff, support tickets, sales calls and documents. Made-up test questions help, but live wording tells us how people actually ask.
For every question we define which passages should come back. If the right evidence does not appear, the answer should not be trusted.
Missing policies, requests for personal data, legal judgement and outdated documents.
We rerun the same questions after document updates and tuning, so your team can see when quality moves.
If HR owns one policy, compliance another and operations keeps a working copy, the assistant needs a rule for conflicts. Otherwise it will confidently cite the wrong version.
Staff, clients and the public usually need different access rules, and an answer should never reveal a source its reader cannot see.
This is where many knowledge base projects fail: the interface looks finished, but nobody has measured whether the right passages come back. We treat answer quality as the product. More on how to judge one in five tests before choosing a model.
07 Data and governance
An assistant is only as clean as the material it reads. PDFs hide tables, scans lose footnotes, folders hold drafts, and old policies sit beside current versions. We look at that mess directly instead of pretending it will sort itself out. Getting documents ready usually means reading scanned files, tagging each one with its owner, date and version, removing duplicates and applying access rules.
Sources are part of the trust. A staff assistant may only need a file name and a paragraph. One that faces clients may need a title, a date, a version and a short note on why that source was used. If your documents change often, the update path is built in: an overnight refresh, a trigger from your document store, an approval queue, or a staged release where new documents are tested before the public assistant sees them.
For personal data we document where it flows, how long it is kept, who can reach it and how it is deleted. We check governance before architecture too. The ICO guidance on AI and data protection, the UK government introduction to AI assurance and the OWASP Top 10 for LLM Applications point to the same need: know what data is used, test the system, document the risk and keep people accountable. Each month we go through the questions it could not answer, the weak sources and the effect on your team. An assistant improves when real use is reviewed. It decays when nobody looks.
08 Patterns we use
Pattern · internal policy
Staff ask in team chat and get a plain-language answer with a link to the exact paragraph. When the policy is silent, the assistant says so and names who to ask. The compliance lead gets an admin view, the library updates as documents change, and accuracy is reviewed against a set of questions we hold back.
Pattern · property knowledge
Each answer tied to portfolio notes, tenancy guidance and approved procedures. Plain questions are searched by meaning, property codes by exact match. Document owners get a review queue, versions are tracked and answer quality is checked weekly.
These describe how we would build each one, not systems we can document yet. The one we can document in full is below.
09 How we quote
Four to six weeks
Eight to twelve weeks
Rolling monthly
Every build gets a fixed quote after the discovery call, once we have seen your documents and what a wrong answer would cost you. Model usage and hosting are passed through at cost and itemised.
10 Where this goes deep
It needs your documents wired in, and its answers are worth most when something acts on them. In lettings, it sits behind the property work.
Agents wired into the document stores, inboxes, CRM and databases the assistant reads from, with access controls kept intact.
Read moreWorkflows that take an answer and triage, file or route the next step, with a person approving anything that matters.
Read moreWhen the answers should sit inside a client portal or an internal tool, that is a custom build.
Read moreEnquiries, viewings, compliance renewals, maintenance triage and landlord reporting for UK letting agencies and property managers.
Read moreNot sure a knowledge base is what you need? Our AI consultancy starts with how the work actually runs, and we will tell you if the answer is to build nothing.
11 Case
Public questions about the Renters' Rights Act, answered only from the legislation text and the firm's own guidance notes. Anything outside it is refused and handed to a person, and every conversation is logged.
CaseRegulated services sectorUK · anonymised
How we built a grounded Q and A assistant that answers only from approved source text, refuses unsupported questions and logs every interaction for review.
Grounded Q and A system case study82%
FAQs auto-resolved
Routine questions answered without staff intervention.
0
Hallucinations
Measured across six months of logged use.
18k
Characters indexed
Approved source text loaded into the knowledge base.
Names redacted out of respect for our clients.
12 Questions
The questions we hear most, with the answers we give on a call.
It is an assistant that answers questions from your own documents, policies and data instead of from general internet knowledge. We set up a private library of your material. When someone asks a question, the assistant looks up the most relevant passages and writes its answer from those, with a link back to the original. If the passages do not cover the question, it says so.
No, and that is deliberate. Training bakes information into a model, where you cannot trace it to a source or cleanly remove it. Our assistant looks your documents up every time it answers, so when a document changes the answers follow, and each answer can show where it came from. Training a model is for changing how it writes, such as tone or format, not for keeping it up to date with your policies.
By default it does not answer. It tells the person your documents do not cover it and passes the question to a human, with the conversation attached. You decide the policy, and for regulated sectors we usually lock it to answers backed by a source. Where you allow broader answers, they are clearly marked.
For a one-off job, such as reading one long contract, that works. For something your team relies on every day it falls short. It cannot promise to answer only from your documents, it gives no dependable source, it does not know who is allowed to see what, and you get no log to review. A knowledge base is built around those four things.
It answers only from the passages it found, checks whether that evidence is enough and refuses when it is not. Before launch we build a test of real questions, the passages that should come back and the questions it must refuse, then rerun it after every change. On the system we run for a regulated services firm, we recorded no made-up answers across six months of logged use, because anything unsupported was refused rather than padded out. That is one system on a small, carefully prepared set of documents, so we treat it as evidence for the method, not a promise for every build.
Before launch we measure whether the right passages come back, how useful the answers are and whether it refuses what it should. After launch we watch how many questions it answers, how many it refuses and how many fewer reach your team. Each month we go through the questions it could not answer and the sources that looked weak, and tune from there.
Built in the UK. Hosted on managed servers in the EU, under UK GDPR. If your policy needs it, the document store, search index, logs and workflow layer can run inside your own environment instead. For sensitive content we default to keeping the source documents inside your environment and indexing only what the assistant needs.
Yes. Staff, clients and the public usually need different access rules, so we tag each source with who may see it, and an answer never reveals a source its reader cannot see. We tag every source by audience from day one, because separating public guidance from internal guidance later is harder.
A focused proof of concept on a single set of documents is usually four to six weeks. A production system with access controls, an audit log and a branded surface (web, team chat, WhatsApp) is typically eight to twelve weeks.
We give you a fixed quote after a discovery call, once we have seen your documents and what a wrong answer would cost you. The running cost depends on volume and on which model the system uses, so we size it to your documents and put it in the proposal before anything is committed. Hosting and model usage are passed through at cost and itemised, so you can see exactly what the system consumes.
We set a refresh schedule that matches the source. Some systems update overnight. Others watch folders or databases and update within minutes. Each answer can show which version of the source it used, so old material is easier to spot.
Yes, but quality varies. Ordinary PDFs are usually straightforward. Scans and images need text recognition and a review step, because tables, stamps and handwritten notes can be misread. We flag low-confidence documents instead of pretending every page was read cleanly.
Yes, if it is designed that way. We map personal data, define the lawful basis and retention, restrict access, log usage, support deletion, and run a data protection impact assessment when the risk calls for one.
We choose the database around your documents, your compliance needs and the running cost. For most UK businesses, Postgres with vector search is simpler and cheaper to operate, and it is our usual choice. If you already run something else that fits, we will say so.
Yes. Engineers call the technique retrieval-augmented generation, or RAG: look up the relevant passages first, then write the answer from them. It is what sits underneath. We lead with what it does for you, answers from your documents with the source shown, because that is what you are buying.
13 Enquiry
Tell us what you would like an assistant to answer from. We come back with first questions or a discovery call slot.