Glossary

Context is everything one answer depends on: the current question, conversation history, user memory, page context, city, locale, site identity and merchant identity. It is passed as one typed envelope into routing and runtime so every entry point behaves consistently.

Merchant Runtime powers merchant and industry agent conversations, handling professional answers, boundary control and lead conversion. It is the only runtime permitted to create leads and booking writes; lifestyle and content runtimes never do.

RAG (Retrieval-Augmented Generation) retrieves relevant knowledge before generating an answer and grounds the answer in what was retrieved. Platform retrieval is script-agnostic across Simplified and Traditional Chinese. RAG only solves finding the basis — whether an answer may be given still depends on fact and skill-pack rules.

Runtime is the execution environment where an agent actually runs: receiving the request, assembling context, invoking the brain and tools, producing the reply and recording events. Runtimes are isolated per product domain, and each domain permits exactly one canonical runtime.

Website Runtime is the execution environment on the distribution side: it resolves site identity, loads site configuration and branding, determines default locale and city, and stamps attribution identifiers on every event. Unknown hosts fall back to a neutral shell with no partner branding and no partner attribution.

Widget Runtime mounts, unmounts and instruments embedded assistants on external pages, turning routing decisions into rendered experiences. It only renders and reports events — it never writes attribution or settlement data directly.