Error Handling

Errors from ML Kit carry a stable code from AndroidAIErrorCodes you can branch on.

Catching Errors

import {
  AndroidAIErrorCodes,
  type AndroidAIError,
  android,
} from '@react-native-ai/android'
import { generateText, streamText } from 'ai'

try {
  await generateText({
    model: android(),
    messages,
  })
} catch (error) {
  const { code } = error as Partial<AndroidAIError>

  if (code === AndroidAIErrorCodes.RateLimited) {
    // Wait a few seconds and retry.
  }

  throw error
}

With streamText, errors arrive as error parts in stream:

const result = streamText({
  model: android(),
  messages,
})

for await (const part of result.stream) {
  if (part.type === 'error') {
    const { code } = part.error as Partial<AndroidAIError>

    if (code === AndroidAIErrorCodes.ModelUnavailable) {
      // Show fallback UI.
    }
  }
}

AICore rejects requests an app sends in quick succession, so space out automated requests by a few seconds.

Error Codes

Code	Meaning
MODEL_UNAVAILABLE	The device can’t run the model, or needs an AICore or system update
MODEL_DOWNLOADING	The model is still downloading
DOWNLOAD_FAILED	Downloading the model failed
FEATURE_NOT_SUPPORTED	ML Kit reports the request needs a feature the model doesn’t have
RATE_LIMITED	AICore is busy or the app used up its quota
BACKGROUND_USE_BLOCKED	AICore only serves apps in the foreground
CONTEXT_WINDOW_EXCEEDED	The request is longer than the context window
INVALID_MESSAGE	The prompt has something Gemini Nano can’t take, such as audio
INVALID_SCHEMA	A tool’s JSON Schema can’t be converted
MAX_TURNS_EXCEEDED	ML Kit’s limit on turns was reached
UNSUPPORTED_PLATFORM	The provider was called on a platform other than Android
GENERATION_ERROR	Any other failure