Prompt API Sampling Parameters

concepts

1. Sampling Playground

   Live sliders for temperature and topK, with a prompt input. Submit and watch the same prompt produce different outputs at different sampling settings. Shows defaults via LanguageModel.params().

2. Determinism Test

   Fix the prompt, pick a temperature, run it N times in a row. Count distinct outputs. The other half of the sampling-parameters story: how to actually get repeatable answers out of the on-device model for tasks that need it.

3. Creative Writing Studio

   Choose a preset (Deterministic / Balanced / Default / Creative / Chaotic) or dial your own temperature and topK. The same creative-writing prompt runs twice in parallel — compare how much the outputs diverge at each setting.

4. Creative Parameter Explorer

   Three side-by-side panels (Deterministic / Balanced / Creative) send the same editable prompt via LanguageModel.create() with different temperature/topK configs. Custom parameter sliders cover temperature, topK, and topP. A Markov chain simulator provides fallback output when the API is unavailable.

5. TopK Visualizer

   Canvas chart showing the token probability distribution as you adjust temperature and topK sliders. See exactly how temperature flattens or sharpens logits before softmax, and how the topK cutoff line eliminates the long tail. Token chips show which tokens survive sampling, with entropy and top-1 probability readouts. Five presets from deterministic to chaotic.

6. Local-First Workflow

   Three real-world app flows — Code Reviewer, Customer Reply, Brainstorm — each with preset parameter recommendations for the task. Dial temperature and topK yourself and see how the model's tone, precision, and diversity change per flow. All on-device.

7. Quality Control Panel

   Run the same prompt through Deterministic, Balanced, and Creative configurations simultaneously via Promise.all(). Compare output diversity, latency bars, and word counts side by side. Choose from four prompt presets (poem, fact, list, story) or write your own.

8. Safety Harness

   12 edge-case tests: API detection, default params read-back, temperature=0 determinism, temperature=2.0 maximum, topK=1 greedy decoding, invalid temperature (negative), invalid topK=0, topK above maxTopK, session reuse, streaming chunks, AbortController cancel, and graceful fallback when the API is absent.

why it exists

Until now developers using the on-device Prompt API had to take whatever sampling behaviour the model shipped with. That meant either consistent-but-boring outputs everywhere, or no way to ask the model to be more deterministic for tasks that need it. temperature and topK are the two standard sampling knobs across most LLM toolkits; exposing them on each LanguageModel instance gives the on-device Prompt API parity with what developers are used to from server-side LLM SDKs.

references

• ChromeStatus entry
• Prompt API spec
• Explainer: sampling parameters