build_index.py

import os
import random
import json
import shutil
import faiss
import numpy as np

from tqdm import tqdm
from FlagEmbedding import FlagLLMModel

def create_index(embeddings: np.ndarray, use_gpu: bool = True):
    index = faiss.IndexFlatIP(len(embeddings[0]))
    embeddings = np.asarray(embeddings, dtype=np.float32)
    # if use_gpu:
    #     co = faiss.GpuMultipleClonerOptions()
    #     co.shard = True
    #     co.useFloat16 = True
    #     index = faiss.index_cpu_to_all_gpus(index, co=co)
    index.add(embeddings)
    return index


if __name__ == '__main__':
    model = FlagLLMModel('/share/chaofan/models/bge-multilingual-gemma2',
                        query_instruction_for_retrieval="Given a question, retrieve passages that answer the question.",
                        query_instruction_format="<instruct>{}\n<query>{}",
                        use_fp16=True)

    avaliable_languages = ['zh', 'en', 'ar', 'bn', 'es', 'fa', 'fi', 'fr', 'hi', 'id', 'ja', 'ko', 'ru', 'sw', 'te', 'th', 'de', 'yo']

    new_dir = '/share/chaofan/code/bge_demo/data'
    new_emb_dir = '/share/chaofan/code/bge_demo/emb'

    for lang in tqdm(avaliable_languages[::-1], desc='language'):
        new_qrels_path = os.path.join(new_dir, lang, 'dev_qrels.jsonl')
        new_queries_path = os.path.join(new_dir, lang, 'dev_queries.jsonl')
        new_corpus_path = os.path.join(new_dir, lang, 'corpus.jsonl')

        os.makedirs(os.path.join(new_emb_dir, lang), exist_ok=True)

        if not os.path.exists(os.path.join(new_emb_dir, lang, 'corpus.npy')):
            data = []
            with open(new_corpus_path) as f:
                for line in f:
                    tmp = json.loads(line)
                    data.append(tmp['title'] + ' ' + tmp['text'])
            
            doc_emb = model.encode_corpus(data, batch_size=256)
            np.save(os.path.join(new_emb_dir, lang, 'corpus.npy'), doc_emb)
        else:
            doc_emb = np.load(os.path.join(new_emb_dir, lang, 'corpus.npy'))
        faiss_index = create_index(doc_emb)
        faiss.write_index(faiss_index, os.path.join(new_emb_dir, lang, 'faiss.index'))