Error Handling

Errors from Core AI carry a code you can branch on. A few errors thrown in JavaScript, such as createTensor rejecting its elements, have none, so check that code is set first.

import { generateText } from 'ai'

try {
  await generateText({ model, prompt })
} catch (error) {
  const { code } = error as { code?: string }

  if (code === 'CORE_AI_MISSING_SPM_PRODUCT') {
    // The task isn't linked: add it to the config plugin's tasks
  }

  throw error
}

With streamText, errors arrive as error parts in fullStream, with the same code.

Error Codes

Code	When
CORE_AI_UNSUPPORTED_OS	The device runs an iOS older than 27
CORE_AI_MISSING_SOURCE	The config has no source, or its bundle:// path is not in the app bundle
CORE_AI_MISSING_SPM_PRODUCT	The call needs a coreai-models product that isn’t linked. The message names the task to add (see Choosing Tasks)
CORE_AI_MODEL_NOT_LOADED	A model handle was unloaded, or never loaded
CORE_AI_SESSION_NOT_FOUND	A native session was closed, or its model unloaded
CORE_AI_UNSUPPORTED_TASK	CoreAI.runTask got a task it doesn’t run
CORE_AI_INVALID_INPUT	An input or option doesn’t match the model or the call. The message says what it takes
CORE_AI_CANCELLED	The abort signal or cancelStream stopped the call
CORE_AI_UNSUPPORTED_CAPABILITY	The model can’t do what the call asks, such as structured output from an engine without guided generation
CORE_AI_ERROR	Core AI or a coreai-models runtime threw the error

Models That Stop Loading

Core AI specializes a model for the device the first time it loads, and caches the result in the app’s Caches/coreai-cache folder. If the app is killed while a model compiles, the cache entry can be left incomplete, and every later load of that model fails with The operation couldn't be completed. (Foundation._GenericObjCError error 0.). Delete the folder, and the model compiles again on its next load:

import { Directory, Paths } from 'expo-file-system'

const cache = new Directory(Paths.cache, 'coreai-cache')
if (cache.exists) cache.delete()