v148 · AI · JavaScript
Local-First Workflow
Three concrete app flows powered by on-device Gemini Nano: a redaction assistant that strips PII, a triage tool that categorises support tickets, and a form assistant that auto-fills replies. All inference runs locally — no server round-trip, no API key.
On-device inference
No API key
No network needed
⚠ Prompt API not detected — these demos use a local simulation so you can explore the workflow. Enable via chrome://flags/#prompt-api-for-gemini-nano or join the origin trial.
Document with PII
Click redacted text to reveal
Redacted output
Press "Redact PII" to process the document…
Category:
Incoming ticket
Classification result
Press "Classify & summarise" to triage the ticket…
Customer message
Tone
Suggested reply
Press "Generate reply" to draft a response…
// Local-first pattern: session + streaming + graceful fallback
async function localInference(systemPrompt, userPrompt) {
if (!window.ai?.languageModel) throw new Error('Prompt API not available');
const session = await window.ai.languageModel.create({
systemPrompt,
temperature: 0.3,
topK: 40,
});
const stream = await session.promptStreaming(userPrompt);
let result = '';
for await (const chunk of stream) {
result = chunk; // cumulative
updateUI(result);
}
return result;
}
see also
- Prompt Playground — direct prompt/response
- Writing Assistant — rewrite modes
- Structured Output Studio — JSON output
- Quality Control Panel — latency & context benchmarks
- Safety Harness — edge cases & error states
implementation reference
Need the exact API surface, compatibility boundaries, errors, lifecycle, and source links? Read the matching gendn reference ↗