Skiing AI Assistant

How this page answers

Ranking questions are answered from the resort index in the browser. Everything else takes two hops: Amazon Nova Pro on Bedrock writes the paragraph, then the browser looks those resort names up in Global Ski Atlas and attaches the links. The model never opens a resort page, and nothing here is a Bedrock Knowledge Base.

1. The browser posts the question. The chat box sends one JSON body, { query }, to API Gateway in us-east-1 (POST /prod/chat). API Gateway invokes a Lambda. The browser does not hold AWS credentials.
2. Lambda puts the question in context and calls Bedrock. The function builds a request for Amazon Nova Pro: a fixed ski-assistant instruction, then your message as the user turn. That instruction is the context. It tells the model to answer as a skiing assistant. It does not include resort records, trail counts, or wiki text. Lambda calls Bedrock with an IAM role. Nova Pro returns plain text.
3. The reply comes back as text. API Gateway returns skiingAnswer. At this point the names are still words. Nova Pro can mention Whistler Blackcomb or a brand like Fischer. It has no page id and no map link.
4. The page queries Global Ski Atlas. While you type, the browser has already loaded /api/wiki/index, the same resort index the rest of the site uses. After the answer arrives, the page scans it for resort names. A name that belongs to exactly one resort becomes a link to /wiki/resort.html?page=…. Longer names win, so “Aspen Snowmass” stays one link. Countries, gear brands, and names shared by two resorts stay plain text. That scan is local. It does not call Bedrock again.

Nova Pro on-demand in us-east-1 is $0.80 per million input tokens and $3.20 per million output tokens. A short question plus the ski instruction is a few hundred input tokens. A reply like the one in the chat is a few hundred output tokens. That is roughly two-tenths of a cent, a fraction of a penny, before the atlas lookup, which is a read of an index this site already serves.