Test chatbot answers and review source citations

Start with current knowledge and a small test set

Before testing, check that your agent exists, the intended source says Indexed for internal retrieval, and Overview reports available AI capacity. A saved or Processed source is not enough. Use non-sensitive examples and make each live request deliberate.

Prepare questions before consuming allowance: one direct question with an explicit answer, one question not covered by the source and one bounded follow-up. You do not need to exhaust the daily allowance to learn whether a source is useful. Current reservations permit roughly four generated answers per owner per day across Playground, widget and API, with further shared/provider limits.

• Prepare focused knowledge for testing
• Understand accepted requests and generated-answer limits

1. Ask a question with an answer you can verify

Open Playground from the workspace navigation. Ask a direct question whose facts appear together in your source. Compare every concrete claim with the current source: conditions, numbers, times and exceptions matter. A fluent response is not proof that the information is correct.

Synthetic example: if your sample FAQ says collection is Monday to Friday, 10:00–17:00, after confirmation, ask “When can I collect an order?” An answer should preserve those conditions. An invented promise of weekend collection is a reason to stop and review the source, not to publish.

2. Review the citation, not just the answer

• Check the displayed reference label and source name. A citation should identify the source used for the relevant business fact.
• Use Knowledge to read the current source text and compare it with the answer. A document citation is not a public download of a private file.
• Webpage citations can link to the original public page. Read the relevant passage; the page’s existence does not prove that it supports every claim.
• Review all parts of a combined answer. One correct citation does not validate an unsupported additional condition or promise.
• If the answer is wrong or the source is stale, correct the authorized business facts, re-index and retest. Public source approval is a separate step after editing.

3. Check an unanswered question without inventing a policy

Use a question the synthetic source does not answer, such as “Is collection available on public holidays?” when no holiday policy is stated. Insufficient knowledge should be treated as a gap for review, not a reason to make up an answer.

An unavailable response can also reflect provider, quota or request-state problems. Check the displayed status and Overview readiness before deciding that the business FAQ is missing. The Opportunities page derives retained insufficient-knowledge website exchanges; it does not include every failed or quota-limited request.

On the website widget, a visitor can choose a consented follow-up when knowledge is insufficient. Playground does not create customer conversations or leads. Saving a lead sends no email or CRM message and is not a promise of a reply.

4. Review context, New chat and uncertain delivery

Playground supports bounded contextual follow-ups using fresh retrieval. Ask a short follow-up only when it has a clear reference to the previous turn. An ambiguous question may require clarification. New chat starts independent context; it does not reset accepted usage or provider allowance.

Playground history stays in the tab and its private server context/retry cache expires after 15 minutes. It is not a durable customer conversation archive. The server-to-server API is single-turn; do not assume the Playground conversation model applies to it.

If delivery is uncertain, use the offered retry for the same request with unchanged text. Do not submit a new operation to bypass an error or quota. Accepted fallbacks and failures can count even when no generated answer is returned.

Decide whether to approve public use

• The source is current, readable and authorized for disclosure, including its name.
• Known answers preserve important exceptions and have an appropriate citation.
• Missing information produces a useful outcome without an invented business claim.
• You understand that private Playground knowledge and published Website sources are separate.
• You have reviewed current capacity and have a human contact path for questions that need it. AI citations do not guarantee accuracy, availability or safety for high-stakes decisions.

• Install and verify the approved website widget
• Read support and unavailable-answer guidance
• Review test context, visitor records and consent