src/embed/hf.ts

import { IEmbeddingFunction } from "chromadb/src/embeddings/IEmbeddingFunction";

// Dynamically import module
let TransformersApi: Promise<any>;

export class TransformersEmbeddingFunction implements IEmbeddingFunction {
  private pipelinePromise: Promise<any> | null;

  /**
   * TransformersEmbeddingFunction constructor.
   * @param options The configuration options.
   * @param options.model The model to use to calculate embeddings. Defaults to 'Xenova/all-MiniLM-L6-v2', which is an ONNX port of `sentence-transformers/all-MiniLM-L6-v2`.
   * @param options.revision The specific model version to use (can be a branch, tag name, or commit id). Defaults to 'main'.
   * @param options.quantized Whether to load the 8-bit quantized version of the model. Defaults to `false`.
   * @param options.progress_callback If specified, this function will be called during model construction, to provide the user with progress updates.
   */
  constructor({
    model = "Xenova/all-MiniLM-L6-v2",
    revision = "main",
    quantized = false,
    progress_callback = null,
  }: {
    model?: string;
    revision?: string;
    quantized?: boolean;
    progress_callback?: Function | null;
  } = {}) {
    try {
      // Since Transformers.js is an ESM package, we use the dynamic `import` syntax instead of `require`.
      // Also, since we use `"module": "commonjs"` in tsconfig.json, we use the following workaround to ensure
      // the dynamic import is not transpiled to a `require` statement.
      // For more information, see https://github.com/microsoft/TypeScript/issues/43329#issuecomment-1008361973
      TransformersApi = Function('return import("@xenova/transformers")')();
    } catch (e) {
      throw new Error(
        "Please install the @xenova/transformers package to use the TransformersEmbeddingFunction, `npm install -S @xenova/transformers`."
      );
    }

    // Store a promise that resolves to the pipeline
    this.pipelinePromise = new Promise(async (resolve, reject) => {
      try {
        const { pipeline } = await TransformersApi;
        resolve(
          await pipeline("feature-extraction", model, {
            quantized,
            revision,
            progress_callback,
          })
        );
      } catch (e) {
        reject(e);
      }
    });
  }

  public async generate(texts: string[]): Promise<number[][]> {
    let pipe = await this.pipelinePromise;
    let output = await pipe(texts, { pooling: "mean", normalize: true });
    return output.tolist();
  }
}