How it works

Structured output usually means: wait for the whole JSON, deserialize, then render. DRYL doesn't wait. A tolerant partial reader turns every chunk into a growing typed snapshot — Value, IsComplete, State and the Raw text — so the surface assembles itself as tokens arrive: image, badges, rating, timeline, donut chart.

Where the innovation lies: progressive materialization of a typed object. You render against snap.Value?.Field and each part appears the moment the model commits it — no spinner-then-snap, no intermediate DTOs.

// Cache the stream in a field so the reference stays stable across renders.
IAsyncEnumerable<string> stream =
    Runner.GenerateStreamingAsync<Recipe>(agent, session, prompt, aiKey: "recipe");

<DrylAiGenerate T="Recipe" Source="stream" Key="recipe">
    <ChildContent Context="snap">
        @if (snap.Value?.Title is { } title) { <h3>@title</h3> }
        @* badges, rating, timeline, donut … each fills in as tokens arrive *@
    </ChildContent>
</DrylAiGenerate>

Throttling the render cadence

A fast model can emit tokens every few milliseconds. Naively that means one Blazor Server re-render — and one full re-parse of the buffer — per token, flooding the circuit. The ThrottleMs parameter caps that: token bursts are coalesced so the surface re-renders at most once per ThrottleMs, while text and lists still grow smoothly.

• Default is 66 (~15fps) — a good balance of liveness and load.
• 0 renders on every token — maximum liveness, maximum cost.
• A slow stream is unaffected: if tokens already arrive slower than the budget, each one still renders. Throttling only ever reduces renders during fast bursts.
• The final snapshot always renders — the tail is never dropped, so the completed object is never truncated.