import os import joblib import numpy as np from app.models import models PCA_PATH = os.getenv( "PCA_PATH", "artifacts/embedding_pca.joblib" ) def generate_embeddings(request): model = models.embedding() vectors = model.encode( request.input, convert_to_numpy=True, normalize_embeddings=False, show_progress_bar=False ).astype(np.float32) # PHASE 1: Embedding generation original_dimensions = vectors.shape[1] # PHASE 2: Dimensionality reduction if request.dimensions: if request.dimensions > original_dimensions: raise ValueError("Requested dimensions exceed model output") if request.dimensions < original_dimensions: if not os.path.exists(PCA_PATH): raise ValueError( "PCA artifact not configured. " "Train and deploy PCA before reducing dimensions." ) pca = joblib.load(PCA_PATH) if pca.n_features_in_ != original_dimensions: raise ValueError("PCA input dimensions mismatch") if pca.n_components_ != request.dimensions: raise ValueError("PCA output dimensions mismatch") vectors = pca.transform(vectors) # Optional normalization after projection if request.normalize: norms = np.linalg.norm( vectors, axis=1, keepdims=True ) vectors = vectors / np.maximum(norms, 1e-12) # PHASE 3: Precision quantization scales = None if request.precision == "float32": output = vectors.astype(np.float32).tolist() elif request.precision == "float16": output = vectors.astype(np.float16).tolist() elif request.precision == "int8": # Symmetric per-vector quantization max_values = np.max( np.abs(vectors), axis=1, keepdims=True ) scale_array = np.maximum( max_values / 127.0, 1e-12 ) quantized = np.clip( np.round(vectors / scale_array), -127, 127 ).astype(np.int8) output = quantized.tolist() scales = scale_array.flatten().tolist() return { "dimensions": vectors.shape[1], "precision": request.precision, "embeddings": output, "scales": scales, "original_dimensions": original_dimensions }