Product rules

What a product rule is

A rule is two things:

1. Keys — the words a rule and a product have in common, usually an ingredient or a product name.
2. The text — the sentence your AI reads word for word.

When a product turns up in an answer, every active rule sharing a key with it — plus any rule you have set to reach every product of the brand — is handed to the AI as binding wording that overrides everything else it found, overrides anything said earlier in the conversation, and overrides what the visitor asks for.

How matching works

A rule reaches the AI in one of two ways, and you pick which one when you write it:

There is no AI in either match, no similarity scoring and no ranking: a key rule is an exact match on a tidied-up version of each key, and a brand-wide rule simply always applies.

Both sides of a key match are tidied identically:

1. everything is lowercased;
2. accents are stripped — Vitamin C, vitamín c and VITAMIN C are one key;
3. punctuation and symbols become spaces;
4. repeated spaces collapse and the ends are trimmed.

So " C-Vitamin, (Ascorbinsav)! " and "c vitamin ascorbinsav" are the same key.

Two smaller consequences: a key made only of punctuation (—, ()) saves happily but can never match anything, and a key-matched rule with no keys cannot be made live at all — if a rule genuinely is meant for everything, set it to every product of the brand rather than leaving its keys empty.

Always, or when relevant

Every rule carries one of two modes, and this is the only control over whether a rule can be filtered out.

When relevant rules are scored against the visitor's question and kept only if they clear a relevance floor. If that scoring ever fails, the system fails in the safe direction: every when-relevant rule passes through, exactly as if it were an Always rule. Nothing is silently withheld.

Sentences your shop sends with a product always behave as Always rules for that product. A document import only ever writes Always rules, consolidated one rule per substance, so a list of permitted wording is in front of your AI every time it talks about that product — not only when a visitor's question happens to match it. That same pattern — Always, one rule per substance — is the one to follow for permitted claims you write by hand, with Copilot or over MCP; When relevant stays yours to choose for everything else.

Writing a rule by hand

Press Add rule on the Rules tab.

Importing rules from a document

If your rules already exist as a document — a regulator's list of permitted claims, an internal wording guide — press Import from document and let the AI turn it into draft rules. Any shape of document works: a claims table, a policy, a contract.

::

This is a hint, never a switch that changes how the file is read. Left on the default, the AI judges from the text itself — and the finished screen tells you what it decided.

Pick the model

The AI model dropdown starts on your organisation's everyday model. The longer or more intricate the document, the more it is worth choosing a stronger one: a dense claims table with footnotes and conditions deserves a top model, a short plain list is fine on a cheaper one.

The size limit follows your choice and is shown right under it — "Gemini 3.1 Flash Lite: up to about 957 pages (670,574 tokens) in one import." If your document is longer than that, split it yourself and import each piece.

Upload it

PDF, Word or plain text, up to 30 MB.

Add instructions (optional)

Anything the AI should know about this particular document — how it is laid out, which parts to ignore, what the keys should be. For example: "The first column is the ingredient; ignore the preamble."

Let it run

The window reports each stage: Mapping the document, then Reading N sections — M done, then Merging and writing drafts, with a running count of the statements found and how many of them the check added.

You can close it — the import keeps running, and coming back to the Rules tab reopens the progress window on the run.

Review the drafts

The finished screen says what it covered: "N draft rules created", "N statements found in the document, M of them added by the check", and how it read the file — "Read as a list of permitted claims · language HU". Every rule arrives as a draft; nothing is live until you approve it. ::

How the document is read

The AI does not blindly chunk your file. It first maps the whole document and divides it into sections along its own content, reads every section separately, and then checks each section again for anything it missed, merged into one sentence, or worded wrongly — that second pass is where the "added by the check" number comes from.

The whole document still has to fit your model's window once, for that first mapping pass. That is what the size limit measures.

What an import produces

• One rule per thing, not one per claim. Every permitted statement about a substance, product or topic is consolidated into a single rule. It opens with the exclusivity sentence — "only the following statements may be made about X, and nothing beyond them" — then lists each statement verbatim and numbered, and states a condition they all share once at the end. A substance with eighteen authorised claims becomes one rule, not eighteen.
• Conditions travel with their claims. Restrictions, conditions and prohibitions about the same thing are folded into its rule; ones that stand on their own become Always rules of their own.
• A rule the document states for everything it covers is written once, as a single rule set to every product of the brand — so it also reaches products the document never names, and products carrying no keys. If the document really governs only part of your catalogue, switch that draft to "products that share a key" while you review it and give it the keys it should have.
• Keys carry synonyms automatically. The import adds the other names for the same thing — common names, abbreviations, chemical forms and salts, spelling variants, and the English name when the document is not in English. A Hungarian rule keyed on Folát also carries Folsav, B9-vitamin and Metilfolát. Rules you type yourself are still yours to spell out, as the warning above says.
• Footnotes that limit a claim — licensed to one company, valid only until a date — go into the note for your team, never into the wording your AI reads.

When an import does not finish cleanly

Two things can happen, and neither loses what was already written:

• Some sections could not be read. The import still finishes, and an amber warning names the parts that were missed. Everything else is kept; import the file again to get the rest.
• The run was interrupted — a deploy or a crash stopped it halfway. The window reads "The import did not finish" and offers Resume, which continues from the sections that never finished and never redoes what was already written. The Rules tab reopens that window on its own when you come back to it.

The approval gate

The Review drafts screen (also reachable from the amber banner above the rules table) is where you go through them:

• Pick an imported document on the left to review just its drafts, or review every draft at once.
• Each row shows its keys, its text, its team note, and a mode dropdown you can change on the spot.
• Edit, Approve or Discard one at a time, or select several and approve or discard them together.
• Discarding shows an Undo in the toast — it is reversible for a moment.

A rule matched by keys cannot be approved without at least one key; approving a batch that contains such rules fails and names them. Rules set to every product of the brand are exempt — carrying no keys is the point of them.

Re-importing the same document

Import an updated version of a file you imported before and the review screen compares the two automatically:

• New — a rule with keys that had none before.
• Changed — same keys, different wording, shown with the old text struck through. Approving it deletes the wording it replaces, so you can never end up with both live.
• Rules that were identical are dropped silently — there is nothing for a person to look at.

Where rules are applied

Every answer a visitor can see:

• the chat widget
• email
• Facebook and Instagram
• AI reply suggestions your agents see in the inbox
• phone calls

They are deliberately not applied to things that never answer a visitor: the inbox panel that shows an agent what the knowledge holds, agent-monitoring checks, and AI social-post generation.

What happens if there are too many rules

The whole block of rules for one answer is capped at 100,000 characters. At the cap, nothing is ever trimmed — instead the lowest-ranked products leave the answer whole, taking their passages, their product cards and their rules with them, until the rest fit in full.

The same applies if the rules cannot be read at all: the products leave the answer. Your AI is never allowed to describe a product while missing some of its rules.

You will not meet this ceiling in normal use — a product's own usage is shown on its page in the catalogue.

Managing rules in bulk

The rules table supports searching the rule text and filtering by status, by mode and by origin (written by hand, from a document, or by an assistant). Select several rows to approve, set to Always, set to When relevant, or delete them together.

A rule that applies to every product of the brand shows an Every product badge where its key chips would be — on the rules table and on the review screen alike.

Copilot can also draft rules for you — see Copilot and MCP. It asks for confirmation before it writes anything, and everything it creates is a draft.