Prompt API

The Prompt API, which calls Gemini Nano in Chrome, has been introduced. It can process natural language requests within the browser, and can be used for purposes such as web page Q&A, classification, filtering, and event extraction.

The main usage flow is to check the model's readiness with LanguageModel.availability(), then open a session with LanguageModel.create() while there is user activation. If a download is required, progress must be detected and communicated to the user.

Development requirements were also specified.

• Supported environments: Windows 10/11, macOS 13+, Linux, or ChromeOS on Chromebook Plus
• Storage: 22GB or more
• Runtime requirements: GPU with more than 4GB VRAM, or CPU with 16GB RAM and 4 or more cores
• Audio input requires a GPU
• Network must be unlimited or not metered

In the API for extensions, params() can be used to check defaultTopK, maxTopK, defaultTemperature, and maxTemperature, and topK, temperature, and signal can be passed to create() to control the session. initialPrompts allows continuing a previous conversation, and adding prefix: true to the last assistant message can guide the output format.

Additionally, expectedInputs and expectedOutputs can be used to specify the input/output formats and languages. Supported languages are en, ja, es, inputs can be text/image/audio, and output can only be text. Examples of multimodal use, such as audio transcription and image description, were also presented.