An owner's Contact record has a "Preferred Contact Language" field set to Spanish. The group wants every agent response for that owner rendered in Spanish without asking the model to re-check the field on every single turn. What is the deterministic way to configure this?
Select an answer to reveal the explanation.
Short Explanation
Read the language once, stash it in a variable, and let every later turn just look at the note instead of re-checking the file. That's the difference between a fact the agent knows and a fact the agent keeps re-guessing.
Full Explanation
Reading the preferred-language value into a variable once at session start, and having every subsequent turn reference that variable, is the deterministic pattern: the value is fetched a single time from the record of truth and then reused reliably, rather than being re-derived by the model on every turn. An instruction asking the agent to "check the field before each reply" is closer to what many specialists first try, but it re-introduces the field lookup, and the associated chance of a missed or misapplied check, into every single turn instead of doing it once and storing the result. Creating a separate topic per language massively over-engineers the problem and conflates topic structure, which exists to route intents, with a presentation preference that has nothing to do with intent classification. Relying on the reasoning engine to detect language from the first message is unreliable and ignores an authoritative field the org already has; a short or ambiguous opening message gives the model little to infer from, whereas the Contact record is the actual source of truth and should be read directly.