Errors

Every error Edgewise throws is an EdgewiseError with a stable code, a message, and usually a hint that says what to do. retryable is true when trying again may work.

import { EdgewiseError, UnsupportedInputError } from 'edgewise';

try {
  await generate({ model: 'text:default', input: [photo, 'What is this?'] });
} catch (err) {
  if (err instanceof UnsupportedInputError) console.log(err.hint); // '"lfm2.5-350m" accepts text. Try "lfm2.5-vl-450m" for image input.'
  if (err instanceof EdgewiseError && err.retryable) retryLater();
}

Class	Code	When
`UnsupportedInputError`	`E_INPUT`	the model does not accept this input type, or a question type
`UnsupportedDeviceError`	`E_UNSUPPORTED`	the device cannot run the model (no WebGPU, browser-only API)
`ModelNotFoundError`	`E_MODEL`	unknown model ID or alias, or a preview model without `allowPreview`
`WrongVerbError`	`E_VERB`	a model used with the wrong verb, such as an embed model in `generate`
`DownloadError`	`E_DOWNLOAD`	a network failure, a bad status, or a SHA-256 mismatch (`retryable` is true except for 4xx statuses)
`StorageQuotaError`	`E_QUOTA`	the browser refused to store the model
`OutOfMemoryError`	`E_OOM`	the GPU or process ran out of memory; try a smaller model or `dtype`
`BackendError`	`E_BACKEND`	ONNX Runtime failed to load or run the model
`SchemaValidationError`	`E_SCHEMA`	the output did not match the schema after a retry; has `raw` and `issues`
`PermissionError`	`E_PERMISSION`	the user blocked the microphone
`AbortError`	`E_ABORT`	the `signal` fired or `cancel()` was called
`ConfigError`	`E_CONFIG`	invalid options

Fallbacks#

configure({
  fallback: {
    onUnsupported: 'next-variant', // or 'throw': try the next device instead of failing
    onBackendError: 'cpu',         // or 'throw': reload on the CPU path when a Transformers.js model fails to load on the GPU
  },
});

info.device in every result tells you whether a fallback happened. With configure({ debug: true }) Edgewise logs each decision.