API reference

The full reference, generated from the source with TypeDoc, is at API reference →.

Entry points#

Import	Contains
`edgewise`	the six verbs, `mic`, `audioSource`, `tool`, question builders, `configure`, `capabilities`, `registry`, `defineModel`, `preload`, `unload`, errors
`edgewise/helpers`	helpers built on the verbs
`edgewise/react`	React hooks
`edgewise/ai-sdk`	Vercel AI SDK providers
`edgewise/test`	mock models
`edgewise/worker`	worker mode: `serveWorker`, `connectWorker`

configure()#

Option	Default
`hub`	`https://huggingface.co`	where model files download from
`cacheDir`	`$EDGEWISE_CACHE` or `~/.cache/edgewise`	servers
`wasmPaths`	jsDelivr	browsers: where ONNX Runtime loads `.wasm` from
`maxLoadedModels`	4	models kept in memory
`memoryBudget`	`'auto'`	bytes loaded models may use together, `'auto'` or `'off'` (Performance)
`preferLowMemory`	`'auto'`	trade speed for memory across devices; on for phones
`allowPreview`	false	allow preview and experimental models
`licenses`	`[]` (all)	licence allow-list
`serverGpu`	`'auto'`	`'auto'`, `'off'` or `'force'`
`fallback`	see Errors
`debug`	false	log decisions

Lifecycle#

import { cache, persist, preload, storage, unload, loadedModels, memoryUsage } from 'edgewise';

await preload(['text:default'], { onProgress });   // download and compile ahead of time
await persist();                                 // browsers: ask to keep the cache when space is low
await storage();                                 // { quota, usage }
await cache.list();                              // cached files
await cache.delete('lfm2.5-350m');               // remove a model's files
await unload();                                  // release all loaded models