"""Frozen text embeddings of every caption (BAAI/bge-small-en-v1.5 via fastembed, ONNX, CPU), computed once and cached. The text embedded is exactly the text BM25 indexes: packet.caption_text(caption) (seven fields joined). Run in an environment with fastembed: python embed.py """ import json, re, sys import numpy as np from fastembed import TextEmbedding FIELDS = ("organ", "location", "lesion_shape_margin", "color", "texture", "distribution", "visible_discriminators") caps = json.load(open(sys.argv[1])) ids = sorted(caps) text = ["; ".join(f"{k}: {caps[i].get(k)}" for k in FIELDS if caps[i].get(k)) for i in ids] m = TextEmbedding("BAAI/bge-small-en-v1.5") E = np.array(list(m.embed(text, batch_size=64)), dtype=np.float32) E /= np.linalg.norm(E, axis=1, keepdims=True) np.savez_compressed(sys.argv[2], ids=np.array(ids), E=E, model="BAAI/bge-small-en-v1.5") print(E.shape)