The words this product is built out of

Answering from a specific source rather than from a model's general knowledge. Here the assistant retrieves a passage from your own pages before it answers, and the answer carries the page it came from. A grounded answer that's wrong is a page you can fix.

Finding the relevant text first and giving it to a model to answer from, rather than asking the model to recall. It's what lets an assistant discuss your refund policy — a model knows nothing about your business, so anything it says unprompted about you is invention.

A confident, fluent, wrong answer. Grounding doesn't make it impossible — nothing does — but citing the source makes it visible and fixable rather than mysterious.

A numeric representation of text that lets passages be matched by meaning rather than keyword. It's why a visitor asking “can I bring my dog” finds a paragraph that says “household pets are welcome”.

A passage a long page is split into — small enough to be a precise answer, large enough to keep its context. Chunking badly is the most common reason a retrieval system gives technically correct but useless answers.

The unit this product is billed in — roughly one exchange, a visitor's question and the answer to it. A conversation is usually several. Telegram alerts never use credits, on any plan.

Passing a conversation to a person. Here it's absolute: once a human takes a conversation over, nothing automated replies into it again, including a scripted flow.

Where a captured lead is sent — Slack, Google Sheets, HubSpot or your own webhook. A lead in a dashboard is one somebody has to remember to look for.

The assistant calling an endpoint mid-conversation to fetch something live — an order status, a booking reference. It happens inside the visitor's turn, so a slow endpoint is a slow answer.

An authored conversation path for the cases that must go a specific way — qualification, intake, triage. Most conversations don't need one; the ones that do usually need to report where people give up.

A standard that lets an AI client read from an external system. This product ships an MCP server, so Claude, ChatGPT or an editor can query your workspace directly.

A database feature that filters rows by a session context inside the engine. It's how workspaces are isolated here — a query that forgets its tenant filter returns nothing rather than everything.

A value you send with a write so that retrying after a timeout returns the original result instead of creating a second record. It's what makes a 201 unambiguous.

A standard error format carrying a machine-readable code and a type URI you can open. Every error this API returns links to a page explaining that exact code.