Retrieval playground
Explore real retrieval over coffee-company docs. No LLM is called.
A direct policy lookup. Inspect the source extract and its lexical score.
Starting the real engine...
document /
No documents cleared the lexical floor. Your application can skip the model call rather than guess.
groundedwork retrieves documents and assembles a prompt. It never calls an LLM or verifies that a source answers the question.
The query-independent system instructions and preferences come first, for cache-aware assembly.
System prefix, retrieved knowledge (if any), then your question. Nothing is sent to a model.
answer is the engine's demo-only ask() response, not generated text. The source extract above comes from the retrieved document body.