edgewise/helpers

Helpers

edgewise/helpers holds one-line shortcuts built on the six verbs. Each is a few lines of code you could write yourself; read the source when you need something slightly different.

import { caption, route, vectorIndex } from 'edgewise/helpers';

Built on generate#

Helper	Does
`transcribe(audio, opts?)`	speech to text with `stt:realtime`
`caption(image, { detail })`	a caption with Florence-2; `detail`: `'short'`, `'detailed'`, `'more-detailed'`
`ocr(image)`	text and regions with Florence-2
`detect(image, find?)`	object boxes, or boxes for a phrase
`watch(video, { prompt, everyMs, onResult })`	ask about a live video every few seconds; returns `stop()`

Built on evaluate#

Helper	Does
`route(text, lanes, { model, threshold, otherwise })`	picks one lane
`classify(text, model)`	the label of a fixed-label classifier
`lint(text, policies)`	`{ passed, findings, answers }`; every policy needs a threshold
`redact(text, { model, types, threshold, mask })`	masks personal data
`replaceSpans(text, spans, mask)`	replaces non-overlapping spans

Built on embed#

Helper	Does
`cosine(a, b)`	cosine similarity
`vectorIndex({ name, model, dimensions, gpu })`	a persistent exact k-NN index with `add`, `remove`, `search`, `query`, `pairs`, `save`, `clear`

Built on forecast and audio#

Helper	Does
`detectAnomalies(series, opts)`	values outside the expected range
`detectSpeech(audio, vadOptions)`	speech segments in a recording, in seconds