Testing with Preview

The Preview tab in the Expert editor lets you test queries against your knowledge base before viewers see it. Refine your system prompt, verify answer quality, and check source references — all without publishing.

What is Preview?

Your internal testing sandbox

Preview is a built-in testing interface within the Expert configuration panel. It uses the same search pipeline as the live Expert — same embeddings, same vector search, same RAG generation — but is only visible to you as the workspace admin.

Think of it as a staging environment for Expert. You can ask questions, see exactly how Expert responds, verify which bites it references, and iterate on your configuration without any risk of viewers seeing incomplete or misconfigured results.

Available on all plans

Preview works on every plan, including Free. You can configure Expert, test queries, and refine your setup regardless of your subscription tier.

Free & Starter plans

Configure Expert and test with Preview. Expert goes live in Hub and Pulse once you reach 30 published bites.

Scale plan

Everything above, plus explicit API keys for external integrations like Zendesk, Intercom, and MCP.

How to use Preview

Preview is accessible from the Expert configuration panel in your dashboard.

1
Open Expert settings
Navigate to the Expert section in your DemoBites dashboard. Select the workspace you want to test.
2
Switch to the Preview tab
Click the Preview tab in the Expert configuration panel. This opens the testing interface.
3
Type a question
Enter a natural language question in the input field — the same kind of question a viewer would ask.
4
Review the response
Expert processes the query and displays the answer, confidence level, source bites with timestamps, and response latency.
5
Iterate
Adjust your system prompt or configuration, then test again. Changes take effect immediately in Preview.

Understanding results

Preview shows everything that the live Expert returns, plus additional diagnostic information to help you evaluate quality.

AI answer

The generated response, written based on your system prompt and the content found in your bites. This is exactly what viewers would see.

Confidence level

"High" means Expert found a strong match. "Low" means it hedged. "None" means no relevant content was found. Use this to identify gaps in your content coverage.

Source bites

The specific bites Expert referenced, with titles, poster thumbnails, and the exact timestamp ranges it drew from.

Response latency

Breakdown of time spent on each pipeline stage: validation, embedding, vector search, and RAG generation. Useful for understanding performance.

Refining your system prompt

The system prompt shapes Expert’s personality and tone. Preview is the best place to iterate on it. Change the prompt, test the same question, and compare how the response changes.

Example system prompts

Product support

“You are a helpful product specialist for Acme Corp. Answer questions about our project management tool. Be concise and link to relevant videos when possible.”

Training academy

“You are a training assistant for new team members. When answering questions, always suggest watching the referenced video for a complete walkthrough. Keep answers encouraging and beginner-friendly.”

Developer documentation

“You are a technical assistant for developers integrating our API. Be precise, use technical language, and reference specific video timestamps for code examples and configuration steps.”

Source display testing

Preview renders sources in the display mode you’ve configured — cards, text links, or thumbnails. Switch between modes and test the same query to see which presentation works best for your content.

Cards — full visual cards with poster images, titles, and links. Best for visually rich content.
Text links — compact clickable titles. Best when answers should be the focus.
Thumbnails — a row of poster images. Best for products where screenshots tell the story.

Tips for effective testing

Get the most out of Preview by testing systematically:

Test common questions first

Start with the questions your support team hears most often. These are the queries real viewers will ask.

Test edge cases

Ask about topics you haven't covered in bites. Verify that Expert returns the fallback message instead of guessing.

Test different phrasings

The same question can be asked many ways. Try "How do I set up billing?" vs "Where is the billing page?" vs "Configure payments" to ensure Expert handles variations.

Check confidence scores

Pay attention to the confidence level. If common questions return "low" confidence, you may need more bites covering that topic.

Iterate on your system prompt

The system prompt has a significant impact on answer tone and structure. Try different approaches and compare results on the same set of test queries.

Tip: Preview works even below the 30-bite threshold. Use it to start shaping your Expert experience while you build up your bite library. By the time you cross 30, your configuration will already be dialed in.