ML Kit GenAI APIs Explained

Building on-device generative features used to mean bundling a model, managing GPU delegates and tuning prompts yourself. ML Kit GenAI packages common tasks as high-level APIs on top of Gemini Nano, which Android's AICore service manages. You call a client, check availability, and get results without network or per-request fees.

The APIs

The shared lifecycle

Example: Rewriting

// implementation("com.google.mlkit:genai-rewriting:<version>")
class ToneRewriter(context: Context) : AutoCloseable {
    private val shorten = Rewriting.getClient(
        RewriterOptions.builder(context)
            .setOutputType(RewriterOptions.OutputType.SHORTEN)
            .setLanguage(RewriterOptions.Language.ENGLISH)
            .build()
    )

    suspend fun isReady(): Boolean = shorten.checkFeatureStatus().await() == FeatureStatus.AVAILABLE

    /** Returns up to a few suggestions; the user picks one. */
    suspend fun shorter(text: String): List<String> =
        shorten.runInference(RewritingRequest.builder(text).build()).await()
            .results.map { it.text }

    override fun close() = shorten.close()
}

Example: Proofreading

// implementation("com.google.mlkit:genai-proofreading:<version>")
val proofreader = Proofreading.getClient(
    ProofreaderOptions.builder(context)
        .setInputType(ProofreaderOptions.InputType.KEYBOARD)     // typed text (vs. VOICE)
        .setLanguage(ProofreaderOptions.Language.ENGLISH)
        .build()
)
suspend fun fix(text: String): String? =
    proofreader.runInference(ProofreadingRequest.builder(text).build()).await()
        .results.firstOrNull()?.text

Example: Image Description

// implementation("com.google.mlkit:genai-image-description:<version>")
val describer = ImageDescription.getClient(ImageDescriberOptions.builder(context).build())

suspend fun altText(bitmap: Bitmap): String =
    describer.runInference(ImageDescriptionRequest.builder(bitmap).build()).await().description

Task API or Prompt API?

• Task APIs (summarise, proofread, rewrite, describe) come with tuned prompts and quality evaluation by Google. Prefer them whenever your feature matches.
• Prompt API gives flexibility for custom tasks such as classifying a message, extracting fields or generating short labels. You own prompt quality and evaluation, and outputs can vary by Gemini Nano version.

Constraints to design around

• Device support is limited. Check status before showing UI, and provide a fallback or hide the feature.
• Foreground only: background inference is blocked.
• Quotas: per-app inference quota (BUSY) and battery-use quotas.
• Input limits and languages vary per API. Summarisation, for example, documents roughly 4,000 input tokens and a small set of languages.

Real-world uses

Common mistakes

• Showing AI buttons without checking feature status.
• Running inference automatically on every keystroke and hitting quotas.
• Applying suggestions automatically without letting the user review them.
• Forgetting close(), which leaks resources.

Frequently asked questions

Conclusion

ML Kit GenAI offers private, on-device generative features with a consistent lifecycle: get a client, check status, download, run inference, close. Prefer the task APIs where they fit, use the Prompt API for custom tasks, and design for limited device support, foreground-only use and quotas.

Next step: build a complete feature with these APIs in How to Build an AI Writing Assistant for Android.