Stuck with carousel + know. base + agent integration
When using the toggle for carousels the agent doesnt produce any. Still behaves as if there was one, though (screenshot)
It should take the data from the knowledge base ( it has, on the website and products it has to show).
What must be done for the agent to show carousels?
2 Replies
Carousels only render when the assistant sends a carousel trace.
Turning ON the “Use carousels”/“Structured response” toggle simply tells the LLM that it is allowed to return that trace type, but the LLM will not invent one for a plain Knowledge-Base answer.
To actually see cards in the chat widget you have to give the agent a way to build the carousel. What to do 1. Make sure you are testing in a Chat interface. Card / Carousel traces are ignored by voice or API-only assistants.
2. Create the trace yourself, either of these ways:
• Add a Carousel step in the workflow – before the LLM response, branch to a Carousel step and hard-wire or variable-fill the cards.
• Generate the trace in a Function / Code step or Custom Action – query the KB (or any API), loop through the returned items and output:
(see “Dynamically render Voiceflow steps” in the docs).
3. If you want the LLM to choose when to send a carousel automatically, add an instruction to your System Prompt such as:
“When you need to recommend multiple products, reply only with a Voiceflow carousel trace containing one card per product (max 5).”
The toggle simply makes that trace type valid; the quality of the prompt determines whether the LLM will output it.
4. Each Card needs, at minimum, a
• Card & Carousel step: https://docs.voiceflow.com/docs/card-and-carousel-step
• Dynamically render steps / traces: https://docs.voiceflow.com/docs/dynamic-rendering
Turning ON the “Use carousels”/“Structured response” toggle simply tells the LLM that it is allowed to return that trace type, but the LLM will not invent one for a plain Knowledge-Base answer.
To actually see cards in the chat widget you have to give the agent a way to build the carousel. What to do 1. Make sure you are testing in a Chat interface. Card / Carousel traces are ignored by voice or API-only assistants.
2. Create the trace yourself, either of these ways:
• Add a Carousel step in the workflow – before the LLM response, branch to a Carousel step and hard-wire or variable-fill the cards.
• Generate the trace in a Function / Code step or Custom Action – query the KB (or any API), loop through the returned items and output:
(see “Dynamically render Voiceflow steps” in the docs).
3. If you want the LLM to choose when to send a carousel automatically, add an instruction to your System Prompt such as:
“When you need to recommend multiple products, reply only with a Voiceflow carousel trace containing one card per product (max 5).”
The toggle simply makes that trace type valid; the quality of the prompt determines whether the LLM will output it.
4. Each Card needs, at minimum, a
title (image, description and buttons are optional). Empty cards will make the widget reserve space but nothing appears—exactly the behaviour you are seeing.
Once the assistant actually returns a type:"carousel" trace, the web-chat widget will render the cards automatically.
Docs for reference• Card & Carousel step: https://docs.voiceflow.com/docs/card-and-carousel-step
• Dynamically render steps / traces: https://docs.voiceflow.com/docs/dynamic-rendering
can agents only produce two cards per carousel?