Skip to main content
Visual Presentations turn your agent into a presenter. Upload a PDF or PowerPoint deck to your library, attach the presentation to an agent, and when a user asks a question that matches the content, the agent shows the relevant slide while it speaks its answer. This turns typical “explain how X works” answers into an actual walkthrough — the agent shows the pricing slide when asked about pricing, the architecture diagram when asked how the system works, and so on.

​
Supported media

Each slide is indexed after upload — the agent uses semantic search to find the right slide based on the user’s question, similar to how Knowledge Base retrieval works.

​
Creating a presentation

1

Open the Presentations tab in Studio

In the Studio’s agent creation flow, open the Visual Presentations tab.
2

Upload your media

Pick a media kind (PDF or PowerPoint), upload the file, and give the presentation a name and optional tags.
3

Wait for processing

Uploads process asynchronously — PDFs are page-split and PowerPoints are slide-indexed. Status moves through pending → processing → completed.
4

Attach to the agent

Once processing is complete, select the presentation in the tab and save the agent. It’s now available for the agent to use in every session.

​
How the agent uses it

During a live session:
  1. The user asks a question.
  2. Before answering, the agent semantically searches the attached presentations for the most relevant slide.
  3. If a strong match is found, the agent both speaks the answer and presents the slide — the widget shows the slide alongside the avatar.
  4. If no strong match is found, the agent falls back to answering from its system prompt and knowledge base as normal — no presentation is shown.
The agent decides when to present. You don’t have to script it.

​
Custom instruction

Each presentation has a custom instruction that tells the LLM how to answer using the retrieved slide. The default is a strict, grounded template:
You can customize this per presentation from the upload modal. Common reasons to change it:
  • Loosen groundedness — allow the agent to elaborate beyond what’s on the slide
  • Change presentation style — force the agent to summarize rather than quote
  • Adjust fallback behavior — change what the agent says when no slide matches
The {query} and {results} placeholders are filled in at runtime.

​
Attaching to an agent

Presentations are attached via the visual_presentations array on the agent config:
Attach as many presentations as makes sense. The agent searches across all of them and picks the best match.

​
Best practices

  • Split large decks by topic. A 100-slide master deck retrieves worse than 5 focused 20-slide decks. The agent picks the right presentation before it picks the right slide.
  • Use clear slide titles. Slide titles are heavily weighted in retrieval. “Q3 Pricing Update” is more findable than “Slide 12.”
  • Prefer PDF over PowerPoint when the source allows. PDFs process faster and produce cleaner slide extractions.
  • Test with real user questions. After attaching, join a test session and ask 5–10 questions the way your users would. Watch which slides come up and which don’t.

​
Presentations vs Knowledge Base

Both are retrieval-based, but they serve different purposes. You can — and often should — use both on the same agent.

​
Next steps

Knowledge Base

Attach text-based sources for grounded written answers.

Prompting

Write a system prompt that plays well with a presenting agent.