Technology

Why a knowledge layer of your own.

Any firm that wants AI to answer from its own knowledge takes one of three paths. Two are quick, and both break in the same place.

1. Upload the documents into a chat tool

   You drop the files into a chat tool, which cuts them into pieces so it can search them. It works for a folder and stops working for a firm. Who wrote a document, when it was valid, which property it belongs to and who may see it: all of that falls away when the file is cut up. Anything that changed yesterday stays wrong until somebody uploads it again.

2. Let every system answer for itself

   Each system answers with the search already built into it: your property system, your drives, your mailboxes. The answer is then only as good as those searches are, and you have no say in what ranks first, what counts as relevant or what happens when nothing matches. A question that reaches across three systems has nowhere to go.

3. Build the knowledge layer yourself

   We connect every source, read it for what it means and hold it in one layer of your own. Relevance, currency and permissions become ours to answer for, instead of whatever each system happens to do. The quality of the search decides the quality of the answer, which is why we do not leave that part to anybody else.

The third path is the one we build.

The layers between a question and an answer.

Five layers. Your data enters at the top, from the systems it already lives in, and never has to move.

Five layers, and a person in the loop

1. Sources, where they already are

   Nothing is moved into a new tool first. Each source is connected where it sits, and read again every time your people add to it. Your property system and your accounting stay the systems of record; the layer reads them, it does not replace them.

2. Capture and distillation

   Every document, email and record is read for what it means rather than for the words in it. A handover protocol is recognized as a handover protocol, and the date in it as a handover date. What comes with the file is kept: who wrote it, when, which property and lease it belongs to, and who inside your firm is allowed to see it.

3. Three searches, at once

   Every way of searching has a blind spot, so three of them run on every question and each one covers for the others. What they are and why is the next section.

4. Fusion and reranking

   Three searches return three lists that cannot be compared to one another. They are merged into one, and what several of them agree on outranks what only one of them was sure about. A document that is right but two years out of date loses to the one written this quarter.

5. The answer, with its sources

   What comes back is an answer with the documents it was built from, not a page of hits where a keyword matched. If the answer is wrong, you can see in one click where it came from. Agents read that same layer the way a colleague does, so an agent works from everything the firm knows and not from the one thread it was handed.

What runs inside those layers, which models do which job and how the pieces are tuned, is not on this page. That part is ours.

Why three searches and not one.

1. Full text, for the exact words

   A clause number, a house number, a file reference, a name spelled the way your colleague spells it. When somebody types the exact string that is in the document, no clever matching should outrank it.

2. Semantic, for the meaning

   The person asking and the person who wrote the answer rarely use the same words. Someone asks about a rent reduction after water damage; the file says the tenant withheld part of the rent after the burst pipe. No word matches, and it is still the right document.

3. Traversal, for the connections

   Some answers are not in any single document. They are in the path between them: from the property to its lease, from the lease to the tenant, from the tenant to the colleague who handled the last complaint. Following that path is also how the brain can tell you who knows about something when no document does.

Then the three lists are fused, reranked and cut to the handful that answer the question. Only those go into the answer.

Nothing becomes true because the system said so.

Real firms contradict themselves. Two documents carry the same date and different numbers. A decision was made in a call and never written down. A figure was corrected once, in one place. No amount of searching fixes that, and a system that quietly picks a winner is worse than one that admits the conflict.

So the brain flags what does not add up, shows both sides with the documents behind them, and waits for a person who knows. That decision goes back into the layer, so the next question already has it, and so the same conflict is never put to a human twice. This is how a shared memory turns into a single source of truth, and why a person sits in the loop from the first day.

What makes this hard.

Cutting some documents up and searching them takes a weekend. These four parts are why a knowledge layer for a real firm takes longer.

1. Carrying the permissions

   Your systems already decide who may see what, and that decision has to survive the trip into the layer, per document and per person. The brain answers inside the permissions of the person asking, from the first step, instead of answering first and filtering afterwards.

2. Staying current

   A firm changes every day. What was signed this morning has to be findable this afternoon, and the version it replaced has to stop winning. An answer that is quietly six months old costs more trust than no answer at all.

3. Contradictions and the loop around them

   Finding the conflict is one problem. Deciding what is worth putting in front of a person, keeping the decision, and making sure it holds the next time is the larger one, and it is what the whole loop above exists for.

4. Tables and documents in one layer

   Most of the answers that matter live half in a document and half in a table: the lease in the drive, the object and the contact in your property system, the figure in accounting. Both have to sit in the same layer and be readable in one question, which is a different problem from either one alone.

Your data stays yours.

1. The brain is built into your firm

   It is your system, on your side, and what it becomes over the years belongs to your firm. Nothing about your properties, your owners or your tenants is used to make anyone else smarter.

2. Delete means delete.

   What the brain stores are plain files. Delete one and it is gone.

3. It can run in your own building

   For firms whose data is not allowed to leave the house, the whole system runs on your own hardware.