The Relevance Gate: Why Your Bot Will Not Invent Products

TL;DR: The worst failure of an AI seller is a confident lie: a product you do not stock, a price you never set. aiSTAFF puts a relevance gate on every search at a cosine threshold near 0.64. If nothing clears the bar, the bot says "we do not carry that" instead of inventing an item.

The expensive failure mode

A general-purpose language model wants to be helpful, which is the problem. Ask it for "a waterproof bluetooth speaker under 80 lari" and, with no grounding, it will happily describe one: a name, a spec, a price, all invented. A customer believes it, asks to buy, and your team is left explaining that the product does not exist. You have now turned a sale into an apology and a damaged trust. For a store, a hallucinated product is worse than a bot that says "I do not know," because the lie travels further before it breaks. Avoiding this is the whole point of grounding the bot in your real catalog, which our AI sales bot does by design.

Grounded, not guessing

aiSTAFF does not answer product questions from a model's general knowledge. It searches your catalog, embedded into a private vector store, and answers only from what it finds. This is retrieval-augmented generation: the bot retrieves real products first, then speaks about those and nothing else. The retrieval itself blends semantic and keyword search, fused so the best match rises, covered in why keyword search fails ecommerce chat. If you are weighing this against retraining a model on your data, the trade-offs are in RAG vs fine-tuning.

Grounding alone is not enough, though. Retrieval always returns something, even for a query with no real match. The top result for "gaming mouse" in a furniture store might be a mouse-shaped cushion at a low score. Return it and the bot looks broken. The relevance gate is what stops that.

How the relevance gate works

Every search produces matches with a similarity score, a cosine value between the query and each product, where higher means closer in meaning. The gate sets a floor near 0.64. The logic is short:

• If the best match scores above the threshold, the product is a genuine fit and the bot shows it.
• If the best match scores below the threshold, nothing in the catalog is close enough, and the bot returns an honest empty: "we do not carry that."

The threshold is tuned to sit between a real match and a weak coincidence. A search for "gaming mouse" in a furniture store scores low against every chair and table, falls under the gate, and comes back as a clear no. A search for "office chair" in the same store scores high and returns cards. The gate is the difference between a bot that admits a limit and a bot that bluffs.

An empty answer is a feature

An honest "we do not carry that" sounds like a loss, but it is the opposite. It keeps the customer's trust, it tells you a gap exists in your catalog, and it routes the shopper toward something you do stock instead of a phantom. The bot can follow an empty with a useful nudge: "We do not stock gaming mice, but here are the office accessories we have." That turns a dead end into a redirect. The selling engine that wraps all of this is in the hub, the AI chatbot that sells your catalog.

The same honesty shows up in the cart and the suggestions. A multi-item basket never gets padded with an invented add-on, covered in the conversational cart, and related-item upsells are pulled from real stock, explained in show me cheap ones. When a match does clear the gate, the customer sees a full card with price and stock, covered in product cards in chat that convert.

The gate holds in any language

Because a Georgian or Russian query is translated for the search, the gate runs on the translated query the same way it runs on English. A Georgian search for an item you do not carry returns the same honest no, not an invented product, so the guardrail does not weaken when the customer switches language, detailed in how a Georgian customer shops your English catalog. Word forms and synonyms are handled before the gate, so a real product is not wrongly rejected over grammar, covered in comfy chair should find your armchairs.

Tuning the threshold for your store

The gate is a dial, not a fixed wall. Set it too low and weak matches slip through, set it too high and real products get rejected. The default near 0.64 is a sensible starting point, and it is tuned per business against the messy queries your customers send: misspellings, vague descriptions, mixed languages. You watch the dead-end rate and the wrong-match rate and adjust until honest empties happen only when the catalog lacks the item. A healthy store keeps that gate tight enough to never lie and loose enough to never hide a product it sells. The broader build is in the ecommerce chatbot guide.

A worked example

An electronics shop gets two messages a minute apart. The first: "do you have a 65-inch OLED TV?" The bot searches, finds two matching models above the threshold, and returns cards with price and stock. The second, from another customer: "do you sell washing machines?" The shop only carries TVs and audio. The search scores every product low, the best match falls under the gate, and the bot replies that it does not stock washing machines, then points to the audio gear it does have. The first customer got a real product, the second got the truth and a redirect, and neither got a fabrication. That is the gate doing its job, and it is why a grounded bot is safe to put in front of buyers.

Related reading

• The AI Chatbot That Sells Your Catalog
• Why Keyword Search Fails Ecommerce Chat
• The Conversational Cart
• Comfy Chair Should Find Your Armchairs

FAQ

What is a relevance gate?

It is a similarity threshold, near 0.64 cosine, on every product search. If the best match scores above it the bot shows the product; if nothing clears it the bot returns an honest empty instead of inventing an item.

Will the chatbot ever make up a product or price?

No. The bot answers only from your embedded catalog and applies the relevance gate, so off-catalog queries come back as "we do not carry that" rather than a fabricated product, spec, or price.

Is an empty answer bad for sales?

It protects trust and flags a real gap in your catalog. The bot can follow an empty with a redirect to items you do stock, turning a dead end into another chance to sell something real.

Does the gate work for Georgian and Russian queries?

Yes. Queries are translated for the search, and the gate runs on the translated query, so the same honest no applies in any language the customer uses.