AI Conversations
Streaming AI responses, conversation starters, and switching between the AI agent and human agents — the core chat experience visitors have with Vex.
Overview
When the widget's brand has a Vex AI agent, visitors can ask questions in natural language and get instant answers. There is nothing to connect: the brand owns both the widget and the AI. This article covers what the AI conversation looks like and feels like from the visitor's side.
Sending a Message
Visitors type their question in the input field at the bottom of the chat and press Enter to send. The message appears as a right-aligned bubble, just like in a regular live chat conversation.
Messages can be up to 20,000 characters long. Visitors can also attach files (images, videos, PDFs) alongside their question — useful for sharing screenshots or documents the AI might need context about.
Streaming Responses
AI responses appear in real time — word by word — as plain text on the page, next to a small avatar and under a Vex · AI assistant name row; there is no bubble around them. This streaming effect makes the conversation feel natural and responsive, similar to watching someone type.
While the AI is thinking:
- Animated loading dots appear to indicate a response is being generated
- The chat auto-scrolls to keep the latest content visible
- If the visitor scrolls up to read older messages, auto-scroll pauses so they aren't interrupted — scrolling back to the bottom re-enables it
Every message Vex sends carries that label, so visitors always know they're reading the AI assistant rather than a person — see Messages & Media for everything that carries it.
Where the Answers Come From
The AI searches the brand's whole body of knowledge — help articles, documentation pages, uploaded files, website pages that were read for you, question-and-answer entries and the product catalogue. For a brand that has both help content and products, the two are searched separately, at the same time, so a large catalogue cannot crowd your help articles out of an answer, or the other way round. There is no per-widget selection to make: everything the brand knows is in scope for every answer.
If the brand's sources are written in more than one language, the search runs in each of them at once, up to three languages at a time, so a brand with an English and a Hungarian knowledge base keeps full recall in both.
When the AI links to one of your own help articles, that link is built at the moment it answers, from the brand's current public web address. Renaming the address does not break the links in content that was indexed before the rename — every citation is rebuilt from the current address each time an answer is written.
Conversation Starters
Conversation starters are suggested topic buttons offered when a conversation starts — in a new chat, or when a visitor comes back three hours or more after their last message with nobody from your team on the conversation. They appear as outlined chips on the left at the end of the thread (under the greeting in a new chat), giving visitors quick entry points for common questions.
When a visitor clicks a starter:
- The button's label appears as the visitor's message (e.g., "Pricing plans")
- A more detailed prompt is sent to the AI behind the scenes (e.g., "What pricing plans do you offer and what's included in each tier?")
- The AI responds based on the detailed prompt
- All starter buttons disappear
Visitors can simply start typing their own message instead — the starters disappear as soon as any message is sent.
A starter can also be set to start a workflow instead of asking the AI. Tapping it looks the same to the visitor — the label appears as their message and the starters disappear — but the workflow answers instead of the AI. If someone from your team is assigned to the conversation, nothing starts: the label simply reaches them as the visitor's message.
Switching to a Human Agent
After a configurable number of messages, a "Talk to the support team" chip appears on the left, under the last reply. Clicking it sends an agent request — the chat state changes to "awaiting" while a human agent is assigned.
During the wait:
- The visitor sees a confirmation that an agent has been requested
- A "Talk to AI Bot" chip replaces it, letting the visitor cancel the request and return to the AI conversation
- Once an agent joins, the conversation continues in the same chat window — no page change or new thread
The agent can see the full AI conversation history, so the visitor doesn't need to repeat themselves.
Automatic Handoff
In some cases, the AI hands off to an agent automatically:
- When the AI determines it can't help based on the conversation
- When no relevant knowledge is found (if configured)
The visitor sees Yaplet's handoff line — "Let me connect you with one of our human agents. 😊" — in their own language before the transition.
Message Types in the Chat
The conversation thread includes several types of entries:
| Type | What It Looks Like |
|---|---|
| Visitor message | A bubble on the right in your brand color, with the visitor's text and any attached files |
| AI response | Plain text on the left next to a small avatar, under a Vex · AI assistant name row, with the time underneath |
| Agent message | A surface card on the left with the agent's photo, presence dot and name |
| Event | A centered pill on a surface for system events like "Agent joined" or "Workflow started" |
Chat History
Visitors returning to a previously active chat see up to their 20 most recent messages from the last 14 days loaded automatically. If the conversation has more than that, a button appears at the top of the thread — "Load more messages", or "Show older messages" when nothing has loaded yet — and each click fetches the next batch.
The chat maintains scroll position when loading history, so visitors stay anchored on the message they were reading instead of being jumped around.