Talkable
AI that answers your customers from your own data.
3 followers
AI that answers your customers from your own data.
3 followers
Your products and services, orders, payments and deliveries in one place, with an AI in front of them that answers customers around the clock.
Free Options
Launch Team
ClarifyThe full-stack AI CRM for modern GTM teams
Promoted
Standard LLM with system prompt already answers customer questions from docs fine. If you do more than wrap API, show how you resolve contradictroy data sources over time. Real proof of product not prompt.
@konstantin_tikhaev
Fair challenge, and for FAQ-from-docs you're right. That's a prompt, not a product.
We're not answering questions about a doc. We take the order and collect the
payment, which is where the wrapper breaks. Concretely:
The model never states a price that reaches a total. It returns one JSON object
per turn against a strict schema (reply, proposed order, what's been collected,
whether a human is needed) and the server re-prices every line from the catalog
row before an order exists. That replaced five code paths that could create an
order with one.
What a column means is stored data, not inference. At import the model reads the
merchant's table once and proposes a role per column (price, stock, name, SKU),
the merchant can correct it, everything downstream reads what's stored. When it's
unsure it returns "none", because a blank the merchant sees gets fixed and a wrong
guess prices an order.
"Over time" is where the ugly ones live: an amendment used to reset an
already-billed order to draft and bill it twice. That's now one transition table
plus an optimistic lock, so an illegal change can't be expressed at all.
Proof: 316 eval conversations, graded by looking the answer up in the live catalog
rather than against a golden answer, because prices change under us. Every bug we
fix becomes a case. Today's: a customer changing the address on an order already
with the courier got "a colleague will be in touch" and no flag was raised, because
the check sat inside order drafting and only fired when the model happened to
propose an order on that turn.
@martik_avagyan fair correction. server side reconcilation turns model into "parser" not authority. solves transactional integrity
@konstantin_tikhaev
"Parser, not authority" is a better phrasing than mine.
Though it only holds where there's an authority to defer to. Money has one: the
catalog row, the zone table, the transition table. Prose doesn't. If a merchant's
uploaded FAQ says free delivery over 50,000 and their zone table says 1,000, the
order total comes out right and the sentence still comes out wrong. Two retrieval
sources, no precedence between them. We catch that with evals and by logging what
the assistant couldn't answer, which is detection, not resolution. Still open.
The part that does generalize is making the thing decline instead of guess. An
unsure column role stores "none". An ambiguous delivery zone matches nothing.
Blank is cheap, wrong is expensive.
@martik_avagyan static prose not safe for policy. rules live in tables, FAQ must pull values dynamic from schema, or strip numeric claims before retreival