CompanyBrain
Overview
The CompanyBrain is the central data foundation for every use of AI in this system. Where the AI settings govern what it works with, this page governs what the AI knows and how it behaves.
The decisive point: there is no second route. Every AI use case — assistant, summaries, suggestions — draws its context through here. There is no hidden instruction somewhere in the program that you cannot see here. What you store here takes effect everywhere; what you do not store here takes effect nowhere.
Seven tabs: Facts · Sources · Instructions · Policies · Test console · Health · Knowledge gaps.
Core tasks
Maintain facts. Facts are curated company truths — company profile, glossary, tone of voice per brand, guidelines. They flow deterministically into every AI context, that is reliably rather than at the mercy of a search. Each fact carries a category, key, title, content, priority and status.
The glossary is the underrated entry here: in-house abbreviations and terms are otherwise misunderstood by a language model without it noticing.
Set instructions. Instructions govern behaviour — how answers are given, what is to be avoided.
Compare and restore versions. Every effective change to a fact or an instruction leaves a version — whether made in the interface, through the API or by an AI command, including a reset to default. The clock icon (Version history) in the row opens the list, newest first, with reason, time and author. A selected version is shown against the current state (field by field, changes highlighted; unchanged fields can be shown). Restore takes over title, content, language, priority, status and validity (for a fact) or text and the active switch (for an instruction); a fact's category and key stay. The state before is saved as a version of its own first — nothing is lost, and the restore appears as a new version "Restored from v…". A row created before this feature gets its previous state as Starting state on its first change. A row without a brand can only be restored by someone holding the permission for all brands.
Define policies per brand. Policies control what a use case may see at all. They apply per brand, because what is permissible can differ between brands.
Check before arming. The test console shows what a use case actually receives as context. That is the only reliable way to judge a change — the effect of an instruction cannot be guessed dependably.
Check health. The Health tab shows the state of the data foundation: how many chunks exist, how many are prepared for semantic search (Embedded) and how many are still waiting (Pending). If there is a backlog, Backfill embeddings processes it in the background. The run stops at the AI budget cap and resumes next time; the button needs the AI gateway to be switched on. Without embeddings the AI only finds literal matches.
Close knowledge gaps. Every AI answer can be rated “thumbs down”. The Knowledge gaps tab groups this open feedback per use case: Detect gaps shows the recurring questions and — with the AI gateway switched on — a draft for a knowledge article. Detecting uses AI budget (one call per group) and therefore only starts on request. Nothing is published automatically: you copy the draft, create the article yourself and mark the group as handled with Article created — or Dismiss it. Both apply to every feedback entry of the group; it does not come back afterwards.
Fields in detail
Company knowledge is what the AI knows about your house: wording conventions, terminology, principles. It flows into every AI answer without anyone typing it into a prompt.
| Field | Required | Values / format | What it does |
|---|---|---|---|
Key factKey | yes | text, max. 150 characters | The entry's identifier. Do not change it once it is referred to. |
Title title | no | text, max. 255 characters | What the entry is called in the list. |
Content content | yes | text | The knowledge itself. Phrased short and unambiguous it works better than a long paragraph — the AI reads it as an instruction, not as inspiration. |
Category category | no | company, glossary, tone, policy, custom | What the entry stands for. Tone shapes the phrasing, policy sets limits ("no price commitments without a quote"), glossary explains terminology. |
Language locale | no | language code, max. 10 characters | Which language the entry applies to. |
Brand brandId | no | one of your brands | For which brand. That way each brand can have its own voice. |
Rank priority | no | number | What counts first when two entries contradict each other. |
Status status | no | active, draft, archived | Only active flows into answers. A draft can be prepared without taking effect. |
Valid from validFrom / Valid until validUntil | no | point in time | Time-limited knowledge — a promotion period, opening hours, a transitional rule. After it passes the AI no longer draws on it instead of repeating outdated commitments. |
Settings & permissions
- Permission
platform.brain.manage. Without it the page is unreachable. - Customised entries are marked. Shipped facts and instructions carry a Customised marker once
you have changed them — and are then not overwritten by updates.
- Permissions still apply. The brain does not widen what anybody may see; it only curates what
the AI receives as context.
- The portal assistant sees less. For customers in the portal, a narrow, fail-closed slice
applies.
FAQ & troubleshooting
The AI answers using the wrong in-house terms. Extend the glossary. Without an entry the model guesses — and guesses plausibly, which makes the error hard to spot.
A change appears to have no effect. Check it in the test console. It shows the actual context, not the expected one.
The assistant mentions data it should not. Check the policy of the use case in question and of the brand.
Where do I set the model? Not here — under AI settings.