Streaming responses to the browser

Time to first token is the number that matters

A complete response can take ten seconds; the first token usually arrives in well under one. Streaming converts a long wait into immediate feedback without making the model any faster.

The provider call stays server-side. The browser talks to your route, your route talks to the model, and the API key never leaves the server — streaming does not change that boundary.

Append to a buffer and render the whole buffer each time. Rendering per chunk causes visible flicker and breaks any markdown that spans chunks.