@eidentic/transformers

Local embedder and reranker for Eidentic — run bge-small-en-v1.5 (384-dim embeddings) and a cross-encoder reranker via @huggingface/transformers, entirely in-process with no API key or external service required. Implements EmbeddingPort and RerankPort from @eidentic/types.

Install

pnpm add @eidentic/transformers @huggingface/transformers

Usage

import { LocalEmbedder, LocalReranker } from "@eidentic/transformers";

// Load once (downloads model on first run, then caches)
const embedder = await LocalEmbedder.load(); // bge-small-en-v1.5, dim=384
const reranker = await LocalReranker.load(); // cross-encoder/ms-marco-MiniLM-L-6-v2

// Single embedding
const vec = await embedder.embed("What is the capital of France?");
console.log(vec.length); // 384

// Batch embedding
const vecs = await embedder.embedBatch(["hello", "world"]);

// Rerank candidates
const reranked = await reranker.rerank("What is Paris?", [
  { id: "a", text: "Paris is the capital of France.", score: 0 },
  { id: "b", text: "London is the capital of England.", score: 0 },
]);
// reranked sorted by relevance score descending

Links

• GitHub
• Issue tracker
• Root README

Apache-2.0