Class TransformersEmbeddingProvider

src/embeddings/transformers.ts:41–154

}

export class TransformersEmbeddingProvider implements EmbeddingProvider {
  readonly name = 'transformers';
  readonly modelName: string;
  readonly dimensions: number;

  private pipeline: FeatureExtractionPipelineType | null = null;
  private ready = false;
  private initPromise: Promise<void> | null = null;

  constructor(modelName: string = DEFAULT_MODEL) {
    this.modelName = modelName;
    this.dimensions = MODEL_CONFIGS[modelName]?.dimensions || 384;
  }

  async initialize(): Promise<void> {
    if (this.ready) return;
    if (this.initPromise) return this.initPromise;

    this.initPromise = this._initialize();
    return this.initPromise;
  }

  private async _initialize(): Promise<void> {
    try {
      if (process.env.CODEBASE_CONTEXT_DEBUG) {
        console.error(`Loading embedding model: ${this.modelName}`);
        console.error('(First run will download the model - this may take a moment)');
      }

      const { pipeline } = await import('@huggingface/transformers');

      // TS2590: pipeline() resolves AllTasks[T] — a union too complex for TSC to represent.
      // Cast to a simpler signature; the actual return type IS FeatureExtractionPipelineType.
      type PipelineFn = (
        task: 'feature-extraction',
        model: string,
        opts: Record<string, unknown>
      ) => Promise<FeatureExtractionPipelineType>;
      this.pipeline = await (pipeline as PipelineFn)('feature-extraction', this.modelName, {
        dtype: 'q8',
        // Limit ONNX Runtime to half cores by default — prevents system freeze during indexing.
        // interOpNumThreads: 1 — no benefit for single-model pipelines.
        // Override via CODEBASE_CONTEXT_ONNX_THREADS env var.
        session_options: {
          intraOpNumThreads: getOnnxThreadCount(),
          interOpNumThreads: 1
        }
      });

      this.ready = true;
      if (process.env.CODEBASE_CONTEXT_DEBUG) {
        console.error(`Model loaded successfully: ${this.modelName}`);
      }
    } catch (error) {
      console.error('Failed to initialize embedding model:', error);
      throw error;
    }
  }