import { pipeline, type FeatureExtractionPipeline } from '@huggingface/transformers';
const MODEL_ID = 'Xenova/all-MiniLM-L6-v2';
export const EMBEDDING_DIMENSIONS = 384;
/**
* Local embedding service using transformers.js.
* Generates 384-dimensional vectors for semantic search.
*/
export class EmbeddingService {
private embedder: FeatureExtractionPipeline | null = null;
private initPromise: Promise<void> | null = null;
/**
* Initialize the embedding model (lazy, called on first use).
* Downloads model on first run (~23MB, cached in ~/.cache/huggingface).
*/
async init(): Promise<void> {
if (this.embedder) return;
if (this.initPromise) return this.initPromise;
this.initPromise = (async () => {
this.embedder = await pipeline('feature-extraction', MODEL_ID, {
dtype: 'fp32',
});
})();
return this.initPromise;
}
/**
* Generate embedding vector for text.
* @returns 384-dimensional normalized vector
*/
async embed(text: string): Promise<number[]> {
if (!this.embedder) {
await this.init();
}
const output = await this.embedder!(text, {
pooling: 'mean',
normalize: true,
});
// Output is a Tensor, convert to array
return Array.from(output.data as Float32Array);
}
/**
* Check if the service is ready (model loaded).
*/
isReady(): boolean {
return this.embedder !== null;
}
}
