Knowledge

Knowledge keeps your company's explanations, ways of working and business assumptions in one place. It comes installed and cannot be removed.

It does three things. You bring in material. You adopt the knowledge the AI finds in it. When you correct something, everything that uses it is marked. Website, Marketing, Sales and Support all work from the knowledge kept here.

Click Knowledge in the menu bar to open the top page, with the cards Knowledge health, Proposals, Quick Links and Meaning Objects. Everyone sees the same menus.

• Business Definition — write down, as statements, what this business is, what it does and what it is deciding
• Meaning Objects — where you read this brand's knowledge. Three tabs — pick from the list, search by meaning, see what is missing — and the record you pick opens on the right
• Inquiries — the first messages that came in; decide whether each becomes a support case or a lead
• Proposals — changes to knowledge the AI suggested; read the reason and the diff, then accept or reject
• Sources — where material comes in, organised into buckets and items
• Extractions — candidate knowledge the AI pulled from your material; keep or discard each one
• Graph — see your knowledge and its links as a picture

The chapters below explain each menu's screen: how to read it and what you can do there.

Business Definition records your business one statement at a time. Each statement has a kind — fact, definition, hypothesis, planned, candidate, constraint or decision point — and you can read them all together under "Our Business (as a document)".

• Draft from a description or URL / Draft from a document — the AI drafts statements in a minute or two; drafts arrive as "not confirmed yet" and nothing is settled until you confirm
• Add a statement / Edit / Add detail — add, correct or expand statements yourself
• Confirm — mark that you have checked it; a hypothesis or plan stays a hypothesis or plan
• Withdraw — take a statement down, with a reason

Each statement shows its state: not decided yet, needs more detail, not confirmed yet, a change is waiting for approval, or confirmed.

Meaning Objects is where you read the knowledge kept for this brand: pick on the left, read on the right. It is the only screen that shows a single piece of knowledge. When you open knowledge from the dashboard, the Graph, approvals or anywhere else, you land here with that record open on the right, and its row is marked in the list on the left.

The left side has three tabs.

• List — browse and pick
• Search by meaning — find knowledge close in meaning even when the words differ (see the chapter below)
• Missing knowledge — demand your knowledge does not cover yet (see the chapter below)

Reading the list

• One row is one piece of knowledge. If it exists in several languages, such as Japanese and English, the versions are shown together on one row
• Only active knowledge is listed. Retired knowledge is not deleted; filter by status to see it
• The columns are Title, Domain (which of the seven areas of meaning it belongs to), Publish state and Updated. The most recently updated come first

Narrowing the list

• All locales — show one language only
• Search by title — find words that appear in the title. To find knowledge by meaning even when the words differ, use the Search by meaning tab
• Status — at first only Active records are listed. Choose Archived, Superseded or Deleted to see knowledge that is no longer in use

The right pane shows the title, labels for domain, publish state and visibility, and three tabs.

• Current value — the summary, body, review status, confidence, last update and semantic tags. For knowledge that exists in several languages, switch language here. When one language has more than one active row, Duplicate rows appears and you choose the one to keep; the others are retired (you see what will change before you confirm)
• History — the values before each change, and the records of confidence checks. When nothing is listed, it says whether nothing has changed or nothing has been recorded yet
• Relations — the knowledge that points to this one and the knowledge this one points to (see the Relations chapter). Click a name to open that knowledge on the right

You do not rewrite knowledge on this screen. Changes to knowledge take effect when you accept them under Proposals.

Inquiries lists conversations that reached the brand's channels (website chat, LINE and the like) and that nobody has taken on yet. Select one on the left to read the messages, oldest first, on the right.

• The columns are Counterpart, Subject, Channel (where it came in) and Last message. The most recent conversation comes first
• Status — at first only Unhandled conversations are listed. Choose Handled to see the ones already routed
• Search by sender — find a conversation by the sender's address

When you have read it, choose one of two buttons in the right pane. Each asks you to confirm before it acts.

• Make a case — open a support case; suited to problems and fault reports
• Follow up as a lead — make it a sales lead. If that person already has a lead, the conversation is attached to it instead of creating a new one

A routed inquiry moves to Handled. You continue it under Cases in Support or Leads in Sales.

Sources has three panes. Pick a bucket (a container for material) on the left and an item in the middle, and the full text appears on the right.

• New bucket — create a container to keep material apart
• Upload a file — put a Word, text or Markdown file (up to 10 MB) into the selected bucket; its text becomes an item and material for extraction
• Paste a transcript / Upload a transcript file — add transcripts of seminars and meetings, or video captions; nothing is transcribed, so no credits are used
• Upload audio / Transcribe audio from a URL — turn a recording into text; one minute of audio uses one credit

A recording is not transcribed until participants' consent is confirmed, and the audio file is not kept after transcription. Use Set aside to take an item you don't need out of the list. It is not deleted, and its previous state is remembered.

From the material you brought in, the AI pulls out candidate knowledge that captures how you do things, and lists it under Extractions. Select one to read the text, the AI's recommendation, where in which source it came from, and the closest existing knowledge. You decide.

• Keep — keep it as new knowledge
• Update the existing knowledge with this text — when similar knowledge already exists, rewrite that one instead of adding another, so nothing is stored twice
• Discard — drop the candidate; the record of your decision stays

Suggestions that would change existing knowledge arrive under Proposals. Read the reason and the diff, then Accept or Reject. You can also tick several and use Accept checked or Reject checked. Suggestions to merge similar topics arrive here too, and they become one only if you accept.

A relation is a link between two pieces of knowledge, with a direction: from → relation → to. There is no screen that lists relations on their own; you see them in three places.

• One record's relations — the Relations tab on the right of Meaning Objects: the knowledge pointing to this one and the knowledge it points to, by kind and by name
• The whole picture — the Graph
• Approving relations the AI proposed — the approval inbox. Only links the machine made on borderline evidence, which need a person's judgment, arrive there

There are exactly seven kinds of meaning link, and no new kind can be invented.

• Extends — From broadens To
• Depends on — From presupposes To
• Exemplifies — From is an example of To
• Contradicts — From and To disagree
• Constrains — From places a limit on To
• Supersedes — From replaces To
• Overrides — From overrides To

Besides these, there are structural links too, such as which concepts an article covers.

The Missing knowledge tab of Meaning Objects lists demand your knowledge does not cover yet (what people search for that no knowledge answers) and what the AI reported as missing while it was working. When the same gap is reported again, no new row is added; its Seen count goes up by one. The higher the count, the more often work is stumbling on it.

• The columns are Title, Kind, Severity, Domain, Seen, Resolution and Created. The most recently reported come first
• Status — at first only Open gaps are listed. You can also choose Resolved, Ignored or Merged
• The search box — find a gap by words in its title

How to work through a gap

1. Select one and read on the right what is missing
2. Fill in the missing knowledge. Bring in material through Sources, and keep the candidates pulled from it under Extractions
3. When it is written, press Resolve

Resolve only marks the gap as filled; it does not write any knowledge, so do step 2 first. If nothing is really missing, press Ignore. Resolved and ignored gaps are kept, and you can see them again by choosing them under Status.

When a person or a workflow checks whether a piece of knowledge holds, a record is kept. It shows on the History tab of that record on the right of Meaning Objects, beside the values before each change (the decision, the confidence before and after, and the notes). Records only arrive when someone validates knowledge, so there may be none yet.

The confidence set there is what you see as the record's Confidence on the Current value tab. For anything marked Needs revision, read it to see what needs fixing.

The Search by meaning tab of Meaning Objects finds knowledge that is close in meaning, even when the words don't match. A query in Japanese may find knowledge written in English.

• Type at least two characters. Searching starts when you pause; there is no search button
• You can type an ordinary sentence instead of keywords (for example, "How many days do I have to return an item?")
• All locales — narrow the results to one language. Unless you narrow them, results span every language
• Up to 20 results are listed, closest first. Each row shows what kind of thing it is (knowledge, an article chapter, a page description and so on) and how close it is, as a percentage. Click a knowledge row to open it on the right
• The search mode is shown above the results. If meaning-based search was unavailable at that moment and another method answered, you can tell here

Search works per brand. If no brand is selected, or your role is not allowed to search, the screen says so.

For Search by meaning and the AI to find knowledge close in meaning, each piece of knowledge needs a marker of its meaning (an embedding). When you save knowledge, making its marker joins a queue and happens shortly after.

Whether that is caught up is one line on the Knowledge health card of the Knowledge top page: Search catch-up pending. It counts what is waiting, what is being made, and knowledge whose marker still reflects the text before your last change. 0 means caught up. If the number cannot be read, it shows "—" rather than 0; when there are more than were counted, it shows "N+".

While the number is high, search and the AI may find knowledge by its earlier text. The record of each job is checked on the operator side.

Knowledge is organised in three layers. A lower layer must never contradict a higher one.

• Platform — the highest rules of the OS. You cannot rewrite them from your side
• App — how each app works (how marketing derives its claims, sales response targets and so on). From your side it acts as advice
• Organization — knowledge that belongs to your company alone (your brand's voice, judgments for your industry, customer segments and so on). This is where the value is, and the layer you grow day to day

The Platform and App layers are the operator's knowledge and do not appear on your brand's screens. Everything listed in Meaning Objects is Organization-layer knowledge.

Graph draws knowledge as points and relations as lines. You can see at a glance which knowledge stands alone and where clusters form.

Narrowing

• Concept lattice only — draw only the meaning links between concepts (the seven kinds under Relations). Without it, the many lines from articles to concepts bury the links between concepts
• The filter box — narrow the points by words such as the title

Drawing limits: the line at the top shows the total number of points and lines and how many are drawn. Normally the whole brand is drawn. Only past a safety limit (5,000 points or 20,000 lines) are the most connected points drawn first, and the screen warns you. Narrow by words, or switch to the concept lattice only, to look at one cluster.

Actions: click a point to open its details on the right, with the relations going in and out and the knowledge at the other end. Open in Meaning Objects opens that knowledge on the right of Meaning Objects, and clicking the other end's name moves to that knowledge. Zoom with the mouse wheel or a pinch; more labels appear as you zoom in.

When knowledge changes (for example, when you accept a proposal), the plans and per-channel drafts that use it are marked for review. Things a machine can check, such as where a link points, are cleared automatically; anything that needs judgment stays marked until you mark it reviewed.

Earlier text is not erased. What it said before a change is under History on the right.

Website builds pages from this knowledge, Marketing plans content from it, and Sales and Support answer deals and questions with it.

A first message under Inquiries becomes a support case with Make a case, or a sales lead with Follow up as a lead. Nothing you bring in is visible to other companies.