Evidence, not assumptions.

01 / The working set. groundedwork tokenizes a question, ranks documents using BM25 keyword matching, and filters the top candidates against a relevance floor. The surviving documents form the working set.

02 / A source to use. “How do I get a refund?” becomes the term “refund”. In the recorded Nimbus corpus, Returns & Refunds scores 3.8605, above the 0.5 floor. The source says: “Not happy with your beans? Return any unopened bag within 30 days for a full refund. Opened bags qualify for store credit. Email support to start a refund.” Retrieved knowledge is assembled with grounding instructions for your model. groundedwork itself never calls an LLM.

03 / The empty set. “What is the capital of France?” produces no hits in these documents. The working set is empty and grounded is false. No knowledge is injected. The engine's abstention response is “I don't have that in my knowledge base, so I won't guess.” This is about the corpus, not whether the question has an answer elsewhere.

04 / Know the boundary. The floor is a lexical gate, not semantic correctness verification. “when does it show up at my door?” incorrectly retrieves Pausing Your Subscription with score 2.9429, above the same floor. Here grounded=true only means a document passed the gate, not that the source answers the question. Review retrieval quality and evaluate your model separately. Retrieve evidence. Keep the right to say no.

Recorded example · Nimbus sample documents. Deployed engine 0.1.0, 24 fictional documents, top_k 3. These are recorded results, not a live engine execution. The animation shows source extracts, not generated answers. Inspect the evidence JSON.