More accurate, professional assistant replies

Manual testing turned up five ways the shopping assistant's replies could fall short, all fixed now, across every engine tier (AISA Extra Light 1.0, AISA Light 1.0, AISA Standard 1.0, AISA Max 1.0).

Prices and sizes in the assistant's text could occasionally disagree with the product card shown right above it. The assistant now only ever states a figure that comes straight from the catalog data behind that card, never a recalled or estimated number.

The assistant sometimes narrated its own retries (a stray "let me try that again" or similar aside) while it refined a search behind the scenes. Retries and refinements now happen silently; the shopper only sees the finished reply.

The assistant could default to English even when the shopper was writing in another language, including when a message was oddly phrased or tried to instruct the assistant directly. It now reliably follows the language the shopper is actually using, and keeps the conversation going in that language rather than switching back to English.

Replies could read as overly casual, with occasional emoji or slangy phrasing. The default tone is now professional and warm (no emoji or slang), unless a site's configured tone of voice asks for a more casual style.

Replies are now more consistently polished: fewer typos, and no more mixing in English phrasing when replying in another language.

This is a wording change to how the assistant is instructed to respond. No dashboard setting changed, and nothing to configure. Existing configuration versions pick it up automatically since it lives in the engine itself.