How Vex finds answers

You do not configure any of this. It is worth understanding anyway, because it explains why an answer came out the way it did, and what to change when it comes out wrong.

A search over everything the brand owns

When a question arrives, Vex runs a search across everything attached to that brand — its knowledge bases, its documentation, its uploaded files, the website pages we have read for you, your question-and-answer pairs and your product catalogue. There is no per-source setting deciding what is searched; if it is on Brand → Knowledge and it is indexed, it is in scope.

In the chat, a brand that has both help content and products gets them searched separately, at the same time. Whichever kind holds the single best match fills its places first — help content up to 8 pieces, products up to 5 — and the other kind takes whatever is left of the 8 places. That way a large catalogue cannot crowd your help articles out of an answer, and the other way round. Phone calls, the inbox's AI Context panel and the other places that run a single search still use one shared search.

The same content backs every place the brand's AI speaks:

• the chat widget
• phone calls taken by its voice agent
• the suggested replies your team sees in the inbox
• answers assembled from your API tools
• the AI social posts it writes

That is why adding one article improves all of them at once, and why a gap shows up everywhere at once too.

One search per topic, and nothing skipped

Every message that is not a plain greeting, thanks or goodbye is searched before Vex answers — a complaint, an offer or a question that is not phrased as a question all get looked up, so Vex never decides from the wording alone that it "does not need" your content.

When one message asks about several things, Vex looks each one up separately (a message asking two things gets two searches, up to four) and merges what it finds. It answers the topics your content covers and says plainly which ones it has nothing on, instead of guessing at the rest. Only when it finds nothing at all does the "when the AI doesn't know the answer" setting on Personality apply.

At the Deep search depth, Vex also searches the main topic with two more wordings of the question — run in parallel and merged — which catches content that is phrased differently. The level is chosen under Brand settings → Advanced; Brand settings describes the two levels.

Several languages at once

Search has two halves: a meaning-based half, which is naturally multilingual, and a keyword half, which has to be told which language's grammar rules to apply.

The keyword half is run in every language the brand's sources are written in, up to three. A brand holding an English knowledge base and a Hungarian one keeps full recall in both at the same time — a Hungarian question still matches the Hungarian articles properly, instead of being parsed with English rules and half-missing.

The brand's own language always goes first, and it is the fallback for anything that carries no language of its own: uploaded files, scraped website pages and Q&A entries.

Links inside answers are built fresh

When Vex points a visitor at one of your articles, it does not repeat a link that was stored when the article was indexed. It looks up the brand's web address at the moment it answers and builds the link then.

The practical consequence: changing your public web address never leaves a stale link buried inside an AI reply. Links people have already saved elsewhere still break — that part is unavoidable, and Your public web address explains it — but nothing Vex says from that point on is wrong.

Catalogue questions do not go through search at all

Some questions are not about a product — they are about your catalogue. "What do you sell?", "What kinds of protein do you have?", "What are your top three vitamins?" name nothing that a search could match.

For those, Vex asks your catalogue directly for its shape instead of searching:

• with no category, it gets your top-level categories with a product count for each;
• with a category, a condition or a sort, it gets the matching products in one order called "most popular" — products with a priority first, in priority order, then the rest by popularity, out-of-stock products last. The first ones are handed over in full, with live prices and stock — at most 5 per answer — together with the total count, and the category's sub-categories are named at the top of the result. When the visitor asks for more, Vex gets the next 5.

Because this is a lookup rather than a search, the answer is built from your real categories and your real order, not from whichever handful of products a search happened to surface. It works on every text channel and on phone calls (where the agent gets the product names only), and it is offered only when the brand actually has an active product.

Filtered and sorted lists

The same lookup can narrow and sort, for one category or for the whole shop. A question with a condition or asking for a particular order goes this way too:

A product whose every version is out of stock always goes to the end of every list, so a "top three" never opens with something nobody can buy.

Prices and stock are read live

Product prices, sale prices and availability are deliberately not stored in the search index. They live on each version of a product and are read from your catalogue at the moment the answer is written, then put into the text handed to the model.

How a product answer is shaped

In a shop with products, Vex receives at most 5 products per answer, each in full, across every search and catalogue lookup of that answer — and it may name or link only those. By default it lays a product answer out like this:

1. One or two sentences about the visitor's need, from general knowledge — what reaching it takes.
2. Up to 5 base products in one to three short titled groups, one or two sentences each, in the words of the product's own description — with no praise of its own such as "excellent" or "perfect choice".
3. Exactly one question that helps the visitor choose.

The whole answer stays under about 900 characters, unless the visitor asks for detail or a comparison. Dosage, allergens and warnings come up only when the visitor asks about them.

When the visitor asks which one is better or what the difference is, Vex puts the measurable facts side by side — both prices, the amounts, the contents — and never names a winner, never gives a reason for a price difference, and never says two products are the same.

What Vex is and is not allowed to say about a product

Whenever an answer contains products, Vex receives a fixed set of conduct rules together with the product data, on that turn only. You do not configure these and you cannot switch them off — they are the platform's floor, and your own product rules and agent instructions sit on top of them.

In plain terms, Vex must:

• Name only products it was actually shown. It may not name a product, or a whole product type, that did not come back in the results.
• Describe a product only from that product's own sheet. No purpose, quality or property the sheet does not state; no explaining one fact with another ("thanks to…", "because…") unless your data says so; no applying one product's figure to a category. If a visitor asks about something the sheet does not cover, Vex never guesses: it answers with what the sheet does say.
• Add no judgment of its own. Vex describes each product with the selling words of that product's own description, and adds no praise or ranking of its own, about one product or about all of them together — no "excellent", "great", "perfect choice" or "the best", no "premium" versus "budget", no "purer" or "higher grade" between your own products.
• Claim popularity only from your own order. A list in the "most popular" order may be presented as your shop's most popular products, in that order — that is also how Vex answers a question for your best, top, best-selling or trending products. Beyond that it says nothing about how well a product sells — no best-seller, most bought, "what people usually choose" — and it never calls a product the best. Highest rated, newest or cheapest only when a list sorted by your ratings, listing dates or prices says so, only for the products on it, and in its order; a highest-rated claim always comes with the rating and the number of reviews. Vex never ranks products itself by a rating it happens to see.
• Compare with measurable facts only. When a visitor asks which is better, what the difference is or why one costs more, Vex puts the measurable facts side by side — both prices, the amounts, the contents — without the quality words of the descriptions ("premium", "superior", "professional", "luxury", "high-quality"), because next to a second product such a word reads as a ranking. It never names a winner, never gives a reason for a price difference, and never says two products are the same.
• Never mention rules or missing data, or what it cannot say or do.
• Say nothing about other shops' or brands' products, and never disparage a product you sell.
• Say "out of stock" once. Vex mentions it one time, offers a product that is in stock, and promises nothing about when it comes back.

Product rules override everything else

If a product in the answer carries product rules, Vex is handed them as binding wording — after everything it found, and outside the limit on how much retrieved content one answer may carry, so an article can never squeeze a rule out.

Vex is told in so many words that these rules override the content it found, override anything said earlier in the conversation, and override what the visitor asks for. Always rules go through every time a matching product comes up; when relevant rules are filtered against the question that surfaced the product. Draft rules never reach it.

A rule of the type Finds products also works the other way round. For a product search in the chat, your find rules are searched by meaning, and a small AI check on your organization's base model chooses the ones that really fit what the visitor wants. A chosen rule brings the products carrying its keys into the answer and the look-alikes leave, before Vex reads anything — and only the chosen find rules travel with the products. In a shop with at least one such rule, Vex looks up goal and need questions in the search first rather than in the category list, because that is where these rules are chosen. This does not happen on phone calls, on catalogue lookups, or in the answer written after a visitor submits a form. The product rules page explains when to use it — and the one kind of rule it must never be used for.

If the rules for one answer would not fit — or there is simply more retrieved content than one answer can carry — whole products leave the answer, passages, product card and rules together, rather than being described with only some of their rules or their passage being cut mid-text.

Swapping content leaves no ghosts

If you take a knowledge base down, swap another one in, or move a source to a different brand, the old content stops answering immediately. It is not remembered, and it does not keep surfacing in replies after it has gone.

Why isn't Vex using this?

When something you added is not showing up in answers, work down this list. The first two columns of Brand → Knowledge answer most cases on their own.