Playground
What it is
The Playground is a private test chat. You ask a question, and Threadline answers it from a knowledge base, exactly as it would for a customer. Next to the answer, it shows what the AI read and how well it matched.
Nothing you do here is sent to any customer.
The problem it solves
Without a test area, the only way to find out how the AI answers is to wait for a real customer to ask. If the answer is wrong, the customer has already seen it.
The Playground lets you:
- Try questions first. Ask what your customers usually ask and read the answers yourself.
- See why an answer was given. You see the exact passages the AI used.
- Find gaps. If nothing was found, you know what content is missing.
- Check a fix. After editing a document, ask again and confirm the answer changed.
How it works
The Playground uses the same steps as a real conversation. It searches the knowledge base you picked, gives the best-matching chunks to the AI, and shows the reply. The only difference is that the answer stays on your screen.
1. Asking a question
- Open Playground in the navigation.
- Under Answering from, choose a knowledge base, or choose an assistant to test its knowledge base, tone and fallback reply together (each assistant also has a Test link that opens the Playground with it already chosen).
- Type your question and press Enter. You can also click one of the example questions.
The answer appears in the chat. If you have no knowledge base yet, the page asks you to create one first.
2. Reading the "Retrieval debug" panel
The panel on the right shows what the AI read before answering. Click any answer in the chat to see its details.
Confidence
A number from 0 to 1 that shows how closely the best chunk matched the question.
| Label | Meaning |
|---|---|
| Strong match | The content clearly covers this question. |
| Partial match | Something related was found. The answer may be incomplete. |
| Weak match | Nothing really fits. The AI will probably say it does not have enough information. |
In real conversations, an answer whose confidence is below your threshold is handed to a person. You set that threshold under Settings → Escalation. See Handing Over to a Person.
Prompt, Completion and Cost
- Prompt is how many tokens the AI was given to read.
- Completion is how many tokens it wrote.
- Cost is the estimated cost of that one answer.
A token is a small piece of text, roughly three quarters of a word.
Retrieved chunks
The list of passages the AI was given, best match first. Each one shows its text and where it came from. Click the link on a chunk to open it in the editor and correct it.
If you see "No chunks were retrieved for this question", your knowledge base has nothing close to that question.
3. A simple way to test before going live
- Write down the ten questions customers ask most.
- Ask each one in the Playground.
- For each wrong or weak answer, open the chunk and fix it, or add a document that covers it.
- Ask again until the answers are right.
- Try the same question in different words. Customers rarely phrase things the way your document does.
Good to know
- Playground questions use your AI allowance, like real answers do.
- Both Admins and Agents can use the Playground.
- The Playground tests one assistant or knowledge base at a time. To test what a channel will say, pick the assistant that channel uses.
Common questions
The answer is "I don't have enough information", with confidence 0. No matching content was found. Add a document that covers the question. If the knowledge base is new, check that its documents show Ready.
The answer is right here but wrong for a customer. Check that the channel's assistant uses the same knowledge base you tested.
Does the Playground remember earlier questions? It is for testing single questions. Refreshing the page clears the chat.