Knowledge grounding

Paste text sources in the console. They are chunked when you save them. At question time the copilot retrieves the most relevant passages and answers from them.

Retrieval is by embedding similarity, and embeddings are configured for the deployment rather than per app. Until that is switched on, retrieval returns nothing and the copilot answers from your actions and your app state alone — it does not fall back to a text search. The console's knowledge page tells you which of the two you are on.

Pipeline

source → chunking → embedding → ready

Each source shows its live status in the console. Editing a source and removing one are not in the console yet; write to us and we will do it.

What grounding buys you

• No invented policies. The model answers refund questions from your refund policy, or says it does not know.
• Grounded answers. Retrieved passages are supplied to the model as context and treated as data, never as instructions. Rendering a visible citation is up to your prompt and your blocks; the SDK does not draw one.
• Tenant isolation. Retrieval is scoped to your app at the database layer. No other tenant's content can leak into your answers.

Testing retrieval

The console's knowledge page includes a retrieval test box: type a question and see which chunks come back. Until embeddings are switched on it runs a plain text match and labels itself as one, so it tells you what is in your sources rather than what the copilot would retrieve. If the right passage does not surface, rewrite that section of your source; short, factual paragraphs retrieve best.