Grounding

Grounded facts reach the model by two routes. Static context is placed in the system prompt: business hours, policy text, product descriptions that rarely change. Dynamic context arrives through retrieval or function calls at the moment it is needed: this caller’s order status, the actual free slots tomorrow, the current balance. Anything that changes belongs in the second category.

Supplying the facts is necessary but not sufficient. The prompt must also instruct the model on what to do when the answer is absent from the grounded material — say it does not know, offer to find out, or transfer — because the default behaviour of a helpful model is to produce a plausible answer from general knowledge, which is exactly the failure grounding was meant to prevent.

Voice makes verification harder and therefore grounding more important. A chat interface can show a source link the user may click. A spoken answer arrives with no provenance, and callers reasonably treat it as authoritative. Where the stakes are high — money, eligibility, medical or legal information — the safer pattern is to have the agent state only what a function returned and escalate everything else.