pip install -U FlagEmbedding

from FlagEmbedding import FlagRerankerreranker = FlagReranker('BAAI/bge-reranker-large', use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation
score = reranker.compute_score(['query', 'passage'])print(score)
scores = reranker.compute_score([['what is panda?', 'hi'], ['what is panda?', 'The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.']])print(scores)

import torchfrom transformers import AutoModelForSequenceClassification, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-large')model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-large')model.eval()
pairs = [['what is panda?', 'hi'], ['what is panda?', 'The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.']]with torch.no_grad():inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)scores = model(**inputs, return_dict=True).logits.view(-1, ).float()    print(scores)

new_rerank_pairs = [[question, answer] for answer in top_n_answers]

from langchain.vectorstores import FAISSfrom langchain_core.documents import Documentfrom langchain.text_splitter import CharacterTextSplitterfrom langchain.embeddings import HuggingFaceEmbeddings
texts = ['哪个快递公司最好？','我该选哪家快递？','哪个快递最快？','哪家快递服务最可靠？','我应该用哪个快递寄包裹？','哪家快递性价比最高？','发货用哪个快递公司比较好？','哪个快递公司收费最合理？','选择哪个快递更安全？','哪个快递公司的客户服务最好？',    '发顺丰快递']
documents = []for idx, text in enumerate(texts):metadata = {"idx": idx}doc = Document(page_content=text, metadata=metadata)documents.append(doc)
text_splitter = CharacterTextSplitter(separator="\n", chunk_size=512)texts = text_splitter.split_documents(documents)
local_model_name = 'shibing624_text2vec-base-chinese'embeddings = HuggingFaceEmbeddings(model_name=local_model_name)
db = FAISS.from_documents(texts, embeddings)faiss_index = "vectors_db/hln_tb_faiss_index"db.save_local(faiss_index)
question = "发什么快递？"answers = db.similarity_search(question, k=3)print(answers)

[Document(page_content='哪个快递最快？', metadata={'idx': 2}), Document(page_content='发货用哪个快递公司比较好？', metadata={'idx': 6}), Document(page_content='我应该用哪个快递寄包裹？', metadata={'idx': 4})]

from FlagEmbedding import FlagReranker
# 构造一个 FlagReranker 实例，设置 use_fp16 为 true 可以加快计算速度reranker = FlagReranker('BAAI/bge-reranker-large', use_fp16=True)
new_rerank_pairs = [[question, answer.page_content] for answer in answers]# 计算多对文本间的相关性评分scores = reranker.compute_score(new_rerank_pairs)print(scores)

{"query": "Five women walk along a beach wearing flip-flops.", "pos": ["Some women with flip-flops on, are walking along the beach"], "neg": ["The 4 women are sitting on the beach.", "There was a reform in 1996.", "She's not going to court to clear her record.", "The man is talking about hawaii.", "A woman is standing outside.", "The battle was over. ", "A group of people plays volleyball."]}{"query": "A woman standing on a high cliff on one leg looking over a river.", "pos": ["A woman is standing on a cliff."], "neg": ["A woman sits on a chair.", "George Bush told the Republicans there was no way he would let them even consider this foolish idea, against his top advisors advice.", "The family was falling apart.", "no one showed up to the meeting", "A boy is sitting outside playing in the sand.", "Ended as soon as I received the wire.", "A child is reading in her bedroom."]}

torchrun --nproc_per_node {number of gpus} \-m FlagEmbedding.reranker.run \--output_dir {path to save model} \--model_name_or_path BAAI/bge-reranker-base \--train_data ./toy_finetune_data.jsonl \--learning_rate 6e-5 \--fp16 \--num_train_epochs 5 \--per_device_train_batch_size {batch size; set 1 for toy data} \--gradient_accumulation_steps 4 \--dataloader_drop_last True \--train_group_size 16 \--max_len 512 \--weight_decay 0.01 \--logging_steps 10