Skip to main content
Sandbox is the one place to try your assistant before you put it in front of customers. Open an assistant and go to the Sandbox tab. It has two views:
  • Chat — talk to your assistant in the real chat widget, exactly as visitors will.
  • Batch tests — build a set of questions, run them all at once, and see each answer graded, with the knowledge sources the assistant used.
Everything in Sandbox is free and never counts against your credits.

​
Chat

The Chat tab embeds the real chat widget — the same surface your visitors use, so forms, theming, file upload, lead capture, and handoff all behave identically here. Because it is the live widget, what you see is exactly what customers get.
  • Reset restarts the conversation so you can try again with a fresh session (handy after changing instructions or knowledge).
  • Open full screen opens the same preview in a new browser tab.
The Chat preview works even while an assistant is disabled or private, so you can try changes before publishing.

​
Batch tests

The Batch tests tab lets you check how your assistant answers many real questions at once and see each answer graded.

​
Building a question bank

Open the Batch tests tab. You can add questions three ways:
1

Add questions manually

Type a question a customer might ask and click Add question.
2

Generate from your knowledge base

Click Generate from knowledge base to create a starter set based on the content your assistant knows. This works even for a brand-new assistant.
3

Generate from past conversations

Click Generate from past conversations to turn real questions your customers have already asked into a test set. If the assistant has no history yet, use one of the other options.

​
Running a test

Click Run test to send every question through your assistant. Runs happen in the background, so you can leave the page and come back. Each answer is graded automatically:
  • Good — relevant, complete, and supported by your content.
  • Acceptable — mostly right, but incomplete or lightly unsupported.
  • Poor — off-topic, incorrect, or unsupported.
Each result also shows the knowledge sources the assistant used to answer, so you can see why it responded the way it did. A Not grounded flag means the answer made specific claims without a supporting source.

​
Seeing the reasoning

Expand Reasoning on any result to see how the assistant arrived at its answer:
  • Tools fired — which tools the assistant used and what it passed them. For a knowledge base search, this is the exact query it ran.
  • Q&A match — when the answer came straight from a Q&A pair, the matched question and its confidence score.
  • Retrieved passages — the knowledge-base chunks the assistant pulled in, each with its document name and relevance score, so you can judge whether it retrieved the right content.
If a result shows “No tools were used — the model answered from its instructions alone,” the assistant answered without consulting your knowledge base. That is expected for small talk, but for a factual question it is a signal to check the answer closely.
You can see the same reasoning live while chatting in the Chat tab — expand Why this answer under any reply to see the query and passages behind it. This preview is only visible to you; visitors using the real widget never see it.

​
Reviewing and overriding grades

You can override any automatic grade with your own judgment using the grade dropdown on each result. Use this to keep track of which answers you have reviewed and to record where the assistant needs better content or instructions.
If an answer is graded Poor or Not grounded, add or improve the relevant knowledge in the Knowledge Base tab, then run the test again.