@reviewpipe/provider-local

On-device sentiment and embeddings for reviewpipe (Tier 0) via transformers.js. No API key, and no network once the models are cached locally.

Optional dependency

@xenova/transformers is an optional peer dependency. Installing this package alone adds no ML weight — the library is loaded lazily (dynamic import()) only when a model is first needed. To actually run models, install it alongside:

pnpm add @reviewpipe/provider-local @xenova/transformers

If it isn't installed, the first classifySentiment/embed call throws a clear error telling you to add it.

Usage

import { Pipeline } from '@reviewpipe/core';
import { CsvAdapter } from '@reviewpipe/adapter-csv';
import { LocalProvider } from '@reviewpipe/provider-local';

const result = await new Pipeline()
  .source(new CsvAdapter())
  .provider(new LocalProvider())
  .run(csvText);

capabilities: { sentiment: true, embeddings: true, summarization: false, customPrompts: false }.

• Sentiment uses a distilled SST-2 model, which is binary — results are positive or negative (never neutral), except empty input, which returns a no-signal neutral without loading a model. score is signed into [-1, 1].
• Embeddings use all-MiniLM-L6-v2 (mean-pooled, normalized), for theme clustering.

Each model pipeline is created once and reused across calls. Override the model ids or inject a custom pipeline factory via the constructor options.

Tests

Unit tests mock the pipeline, so they neither download models nor require @xenova/transformers — they run in normal CI.

The integration tests download and run the real models and are excluded from the default test run. To run them:

pnpm add -Dw @xenova/transformers   # once, at the workspace root
pnpm test:integration