Prompt API

concepts

1. Prompt Playground

   A chat-style interface that sends your prompts to Gemini Nano via LanguageModel.create() and streams the response back. Falls back to a simulation when the model isn't available, with step-by-step setup instructions.

2. Structured Output Studio

   Force the model to respond in strict JSON using responseConstraint: { type: "json_schema", ... }. Enter a piece of text and see it analyzed and returned as a structured object — sentiment, keywords, a summary — guaranteed valid JSON.

3. Conversation History

   Multi-turn chat backed by Chrome's on-device Gemini Nano via the Prompt API. Watch the context window fill as the conversation grows — when it nears the limit, the oldest turns are pruned and the session is recreated with trimmed history.

4. Function Calling Demo

   Three clickable tool cards (get_weather, set_reminder, search_web) with JSON schemas. Enter a natural language query and watch a 5-step pipeline animate: system prompt engineering → model JSON tool call → parsed tool call → tool execution result → follow-up answer. Heuristic simulation fallback when LanguageModel is unavailable.

5. Writing Assistant

   Five writing modes — continue, improve clarity, summarise, make formal, make casual — backed by session.promptStreaming(). Suggestions stream token-by-token and can be accepted or dismissed. Falls back to a local template simulation when the Prompt API is not available.

6. Local-First Workflow

   Three concrete on-device app flows: a PII redaction assistant, a support-ticket triage and classifier, and a customer-reply form assistant with tone control. All inference runs locally with no server round-trip.

7. Quality Control Panel

   Benchmark the Prompt API across six test cases — short completions, structured JSON output, summarisation, classification, creative writing, and long-context reasoning. Tune temperature and top-K and compare first-token latency vs. total response time.

8. Safety Harness

   14-case edge-case test suite covering API availability, empty input, very long input, prompt injection attempts, structured output validation, session reuse, context overflow, abort/cancel, invalid parameters, and graceful fallback patterns.

why it shipped

Server-side AI inference requires an API key, a network request, latency, and ongoing per-token costs. For use cases that don't need the full power of cloud models — local spell-check, content summarization, form auto-fill, privacy-sensitive user data analysis — a local model is strictly better. Chrome ships with Gemini Nano, a small language model optimized for device execution. The Prompt API exposes it through a standardized JavaScript interface: check availability, create a session with optional system prompt and temperature, send prompts, stream responses. The API is designed to match the emerging W3C Web AI specification so that other browsers can implement the same interface.

references

• ChromeStatus entry
• Chrome Developers — The Prompt API
• New in Chrome 148
• Get started with built-in AI