/** * In-process embedder using @huggingface/transformers (ONNX Runtime). * Default provider for ALL desktop users — zero config, works offline. * Model: Xenova/all-MiniLM-L6-v2 (384 native dims, normalized to target dims). * Downloads ~90MB fp32 model on first use, cached in /models/. */ import path from 'node:path'; import { defaultDataDir } from '../config.js'; import { createLazyEmbedder, type Embedder } from './embeddings.js'; import { createCoreLogger } from '../logger.js'; import { withTransformersModelLoad } from './transformers-model-load.js'; const log = createCoreLogger('inprocess-embedder'); const DEFAULT_MODEL = 'Xenova/all-MiniLM-L6-v2'; export interface InProcessEmbedderConfig { model?: string; cacheDir?: string; targetDimensions?: number; } /** Normalize embedding dimensions: zero-pad shorter, truncate longer. */ export function normalizeDimensions(embedding: Float32Array, targetDims: number): Float32Array { if (embedding.length === targetDims) return embedding; const result = new Float32Array(targetDims); const copyLen = Math.min(embedding.length, targetDims); result.set(embedding.subarray(0, copyLen)); return result; } export async function createInProcessEmbedder(config?: Partial): Promise { const model = config?.model ?? DEFAULT_MODEL; const cacheDir = config?.cacheDir ?? path.join(defaultDataDir(), 'models'); const targetDims = config?.targetDimensions ?? 1024; log.info(`Loading in-process embedding model: ${model} (~90MB fp32 first download)`); const { pipeline } = await import('@huggingface/transformers'); const extractor = await withTransformersModelLoad({ cacheDir, model, load: (canonicalCacheDir) => pipeline('feature-extraction', model, { dtype: 'fp32', cache_dir: canonicalCacheDir, }), onQuarantine: () => log.warn(`Quarantined corrupt embedding model cache: ${model}`), }); const nativeDims = 384; // all-MiniLM-L6-v2 output dimensions log.info(`In-process embedder ready (${nativeDims} native dims → ${targetDims} normalized)`); return { dimensions: targetDims, async embed(text: string): Promise { const result = await extractor(text, { pooling: 'mean', normalize: true }); const raw = new Float32Array(result.data as Float32Array); return normalizeDimensions(raw, targetDims); }, async embedBatch(texts: string[]): Promise { if (texts.length === 0) return []; const results: Float32Array[] = []; // Process one at a time to avoid memory issues with large batches for (const text of texts) { const result = await extractor(text, { pooling: 'mean', normalize: true }); const raw = new Float32Array(result.data as Float32Array); results.push(normalizeDimensions(raw, targetDims)); } return results; }, }; } /** Memory vectors keep all-MiniLM-L6-v2's own size, unpadded (ADR-0007). */ export const MEMORY_EMBEDDING_DIMENSIONS = 384; /** * The embedder memory uses: the in-process model, loaded on first use. Every * process that opens a data directory's memory must use the same one. */ export function createMemoryEmbedder(options: { cacheDir: string; model?: string }): Embedder & { id: string } { const model = options.model ?? DEFAULT_MODEL; const embedder = createLazyEmbedder( () => createInProcessEmbedder({ model, cacheDir: options.cacheDir, targetDimensions: MEMORY_EMBEDDING_DIMENSIONS, }), MEMORY_EMBEDDING_DIMENSIONS, ); return { ...embedder, id: `inprocess:${model}:${MEMORY_EMBEDDING_DIMENSIONS}` }; }