English | 简体中文

Bilingual and Crosslingual Embedding (BCEmbedding) in English and Chinese, developed by NetEase Youdao, encompasses EmbeddingModel and RerankerModel. The EmbeddingModel specializes in generating semantic vectors, playing a crucial role in semantic search and question-answering, and the RerankerModel excels at refining search results and ranking tasks.

BCEmbedding serves as the cornerstone of Youdao's Retrieval Augmented Generation (RAG) implementation, notably QAnything [github], an open-source implementation widely integrated in various Youdao products like Youdao Speed Reading and Youdao Translation.

Distinguished for its bilingual and crosslingual proficiency, BCEmbedding excels in bridging Chinese and English linguistic gaps, which achieves

• A high performance on Semantic Representation Evaluations in MTEB;
• A new benchmark in the realm of RAG Evaluations in LlamaIndex.

Provide a bilingual and crosslingual two-stage retrieval model repository for the RAG community, which can be used directly without finetuning, including EmbeddingModel and RerankerModel:

• One Model: EmbeddingModel handle bilingual and crosslingual retrieval task in English and Chinese. RerankerModel supports English, Chinese, Japanese and Korean.
• One Model: Cover common business application scenarios with RAG optimization. e.g. Education, Medical Scenario, Law, Finance, Literature, FAQ, Textbook, Wikipedia, General Conversation.
• Easy to Integrate: We provide API in BCEmbedding for LlamaIndex and LangChain integrations.
• Others Points:
  • RerankerModel supports long passages (more than 512 tokens, less than 32k tokens) reranking;
  • RerankerModel provides meaningful relevance score that helps to remove passages with low quality.
  • EmbeddingModel does not need specific instructions.

• RAG applications: QAnything, HuixiangDou, ChatPDF.
• Efficient inference: ChatLLM.cpp, Xinference, mindnlp (Huawei GPU).

Existing embedding models often encounter performance challenges in bilingual and crosslingual scenarios, particularly in Chinese, English and their crosslingual tasks. BCEmbedding, leveraging the strength of Youdao's translation engine, excels in delivering superior performance across monolingual, bilingual, and crosslingual settings.

EmbeddingModel supports Chinese (ch) and English (en) (more languages support will come soon), while RerankerModel supports Chinese (ch), English (en), Japanese (ja) and Korean (ko).

• Bilingual and Crosslingual Proficiency: Powered by Youdao's translation engine, excelling in Chinese, English and their crosslingual retrieval task, with upcoming support for additional languages.
• RAG-Optimized: Tailored for diverse RAG tasks including translation, summarization, and question answering, ensuring accurate query understanding. See RAG Evaluations in LlamaIndex.
• Efficient and Precise Retrieval: Dual-encoder for efficient retrieval of EmbeddingModel in first stage, and cross-encoder of RerankerModel for enhanced precision and deeper semantic analysis in second stage.
• Broad Domain Adaptability: Trained on diverse datasets for superior performance across various fields.
• User-Friendly Design: Instruction-free, versatile use for multiple tasks without specifying query instruction for each task.
• Meaningful Reranking Scores: RerankerModel provides relevant scores to improve result quality and optimize large language model performance.
• Proven in Production: Successfully implemented and validated in Youdao's products.

• 2024-02-04: Technical Blog - See 为RAG而生-BCEmbedding技术报告.
• 2024-01-16: LangChain and LlamaIndex Integrations - See more.
• 2024-01-03: Model Releases - bce-embedding-base_v1 and bce-reranker-base_v1 are available.
• 2024-01-03: Eval Datasets [CrosslingualMultiDomainsDataset] - Evaluate the performance of RAG, using LlamaIndex.
• 2024-01-03: Eval Datasets [Details] - Evaluate the performance of crosslingual semantic representation, using MTEB.

First, create a conda environment and activate it.

conda create --name bce python=3.10 -y
conda activate bce

Then install BCEmbedding for minimal installation (To avoid cuda version conflicting, you should install torch that is compatible to your system cuda version manually first):

pip install BCEmbedding==0.1.4

Or install from source (recommended):

git clone [email protected]:netease-youdao/BCEmbedding.git
cd BCEmbedding
pip install -v -e .

Use EmbeddingModel, and cls pooler is default.

from BCEmbedding import EmbeddingModel

# list of sentences
sentences = ['sentence_0', 'sentence_1']

# init embedding model
model = EmbeddingModel(model_name_or_path="maidalun1020/bce-embedding-base_v1")

# extract embeddings
embeddings = model.encode(sentences)

Use RerankerModel to calculate relevant scores and rerank:

from BCEmbedding import RerankerModel

# your query and corresponding passages
query = 'input_query'
passages = ['passage_0', 'passage_1']

# construct sentence pairs
sentence_pairs = [[query, passage] for passage in passages]

# init reranker model
model = RerankerModel(model_name_or_path="maidalun1020/bce-reranker-base_v1")

# method 0: calculate scores of sentence pairs
scores = model.compute_score(sentence_pairs)

# method 1: rerank passages
rerank_results = model.rerank(query, passages)

NOTE:

• In RerankerModel.rerank method, we provide an advanced preproccess that we use in production for making sentence_pairs, when "passages" are very long.

For EmbeddingModel:

from transformers import AutoModel, AutoTokenizer

# list of sentences
sentences = ['sentence_0', 'sentence_1']

# init model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('maidalun1020/bce-embedding-base_v1')
model = AutoModel.from_pretrained('maidalun1020/bce-embedding-base_v1')

device = 'cuda'  # if no GPU, set "cpu"
model.to(device)

# get inputs
inputs = tokenizer(sentences, padding=True, truncation=True, max_length=512, return_tensors="pt")
inputs_on_device = {k: v.to(device) for k, v in inputs.items()}

# get embeddings
outputs = model(**inputs_on_device, return_dict=True)
embeddings = outputs.last_hidden_state[:, 0]  # cls pooler
embeddings = embeddings / embeddings.norm(dim=1, keepdim=True)  # normalize

For RerankerModel:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# init model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('maidalun1020/bce-reranker-base_v1')
model = AutoModelForSequenceClassification.from_pretrained('maidalun1020/bce-reranker-base_v1')

device = 'cuda'  # if no GPU, set "cpu"
model.to(device)

# get inputs
inputs = tokenizer(sentence_pairs, padding=True, truncation=True, max_length=512, return_tensors="pt")
inputs_on_device = {k: v.to(device) for k, v in inputs.items()}

# calculate scores
scores = model(**inputs_on_device, return_dict=True).logits.view(-1,).float()
scores = torch.sigmoid(scores)

For EmbeddingModel:

from sentence_transformers import SentenceTransformer

# list of sentences
sentences = ['sentence_0', 'sentence_1', ...]

# init embedding model
## New update for sentence-trnasformers. So clean up your "`SENTENCE_TRANSFORMERS_HOME`/maidalun1020_bce-embedding-base_v1" or "～/.cache/torch/sentence_transformers/maidalun1020_bce-embedding-base_v1" first for downloading new version.
model = SentenceTransformer("maidalun1020/bce-embedding-base_v1")

# extract embeddings
embeddings = model.encode(sentences, normalize_embeddings=True)

For RerankerModel:

from sentence_transformers import CrossEncoder

# init reranker model
model = CrossEncoder('maidalun1020/bce-reranker-base_v1', max_length=512)

# calculate scores of sentence pairs
scores = model.predict(sentence_pairs)

We provide BCERerank in BCEmbedding.tools.langchain that inherits the advanced preproc tokenization of RerankerModel.

• Install langchain first

pip install langchain==0.1.0
pip install langchain-community==0.0.9
pip install langchain-core==0.1.7
pip install langsmith==0.0.77

• Demo

# We provide the advanced preproc tokenization for reranking.
from BCEmbedding.tools.langchain import BCERerank

from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import PyPDFLoader
from langchain_community.vectorstores import FAISS

from langchain.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores.utils import DistanceStrategy
from langchain.retrievers import ContextualCompressionRetriever


# init embedding model
embedding_model_name = 'maidalun1020/bce-embedding-base_v1'
embedding_model_kwargs = {'device': 'cuda:0'}
embedding_encode_kwargs = {'batch_size': 32, 'normalize_embeddings': True, 'show_progress_bar': False}

embed_model = HuggingFaceEmbeddings(
  model_name=embedding_model_name,
  model_kwargs=embedding_model_kwargs,
  encode_kwargs=embedding_encode_kwargs
)

reranker_args = {'model': 'maidalun1020/bce-reranker-base_v1', 'top_n': 5, 'device': 'cuda:1'}
reranker = BCERerank(**reranker_args)

# init documents
documents = PyPDFLoader("BCEmbedding/tools/eval_rag/eval_pdfs/Comp_en_llama2.pdf").load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1500, chunk_overlap=200)
texts = text_splitter.split_documents(documents)

# example 1. retrieval with embedding and reranker
retriever = FAISS.from_documents(texts, embed_model, distance_strategy=DistanceStrategy.MAX_INNER_PRODUCT).as_retriever(search_type="similarity", search_kwargs={"score_threshold": 0.3, "k": 10})

compression_retriever = ContextualCompressionRetriever(
    base_compressor=reranker, base_retriever=retriever
)

response = compression_retriever.get_relevant_documents("What is Llama 2?")

We provide BCERerank in BCEmbedding.tools.llama_index that inherits the advanced preproc tokenization of RerankerModel.

• Install llama_index first

pip install llama-index==0.9.42.post2

• Demo

# We provide the advanced preproc tokenization for reranking.
from BCEmbedding.tools.llama_index import BCERerank

import os
from llama_index.embeddings import HuggingFaceEmbedding
from llama_index import VectorStoreIndex, ServiceContext, SimpleDirectoryReader
from llama_index.node_parser import SimpleNodeParser
from llama_index.llms import OpenAI
from llama_index.retrievers import VectorIndexRetriever

# init embedding model and reranker model
embed_args = {'model_name': 'maidalun1020/bce-embedding-base_v1', 'max_length': 512, 'embed_batch_size': 32, 'device': 'cuda:0'}
embed_model = HuggingFaceEmbedding(**embed_args)

reranker_args = {'model': 'maidalun1020/bce-reranker-base_v1', 'top_n': 5, 'device': 'cuda:1'}
reranker_model = BCERerank(**reranker_args)

# example #1. extract embeddings
query = 'apples'
passages = [
        'I like apples', 
        'I like oranges', 
        'Apples and oranges are fruits'
    ]
query_embedding = embed_model.get_query_embedding(query)
passages_embeddings = embed_model.get_text_embedding_batch(passages)

# example #2. rag example
llm = OpenAI(model='gpt-3.5-turbo-0613', api_key=os.environ.get('OPENAI_API_KEY'), api_base=os.environ.get('OPENAI_BASE_URL'))
service_context = ServiceContext.from_defaults(llm=llm, embed_model=embed_model)

documents = SimpleDirectoryReader(input_files=["BCEmbedding/tools/eval_rag/eval_pdfs/Comp_en_llama2.pdf"]).load_data()
node_parser = SimpleNodeParser.from_defaults(chunk_size=400, chunk_overlap=80)
nodes = node_parser.get_nodes_from_documents(documents[0:36])
index = VectorStoreIndex(nodes, service_context=service_context)

query = "What is Llama 2?"

# example #2.1. retrieval with EmbeddingModel and RerankerModel
vector_retriever = VectorIndexRetriever(index=index, similarity_top_k=10, service_context=service_context)
retrieval_by_embedding = vector_retriever.retrieve(query)
retrieval_by_reranker = reranker_model.postprocess_nodes(retrieval_by_embedding, query_str=query)

# example #2.2. query with EmbeddingModel and RerankerModel
query_engine = index.as_query_engine(node_postprocessors=[reranker_model])
query_response = query_engine.query(query)

We provide evaluation tools for embedding and reranker models, based on MTEB and C_MTEB.

First, install MTEB:

pip install mteb==1.1.1

Just run following cmd to evaluate your_embedding_model (e.g. maidalun1020/bce-embedding-base_v1) in bilingual and crosslingual settings (e.g. ["en", "zh", "en-zh", "zh-en"]).

python BCEmbedding/tools/eval_mteb/eval_embedding_mteb.py --model_name_or_path maidalun1020/bce-embedding-base_v1 --pooler cls

The total evaluation tasks contain 114 datasets of "Retrieval", "STS", "PairClassification", "Classification", "Reranking" and "Clustering".

NOTE:

• All models are evaluated in their recommended pooling method (pooler).
  • mean pooler: "jina-embeddings-v2-base-en", "m3e-base", "m3e-large", "e5-large-v2", "multilingual-e5-base", "multilingual-e5-large" and "gte-large".
  • cls pooler: Other models.
• "jina-embeddings-v2-base-en" model should be loaded with trust_remote_code.

python BCEmbedding/tools/eval_mteb/eval_embedding_mteb.py --model_name_or_path {mean_pooler_models} --pooler mean

python BCEmbedding/tools/eval_mteb/eval_embedding_mteb.py --model_name_or_path jinaai/jina-embeddings-v2-base-en --pooler mean --trust_remote_code

Run following cmd to evaluate your_reranker_model (e.g. "maidalun1020/bce-reranker-base_v1") in bilingual and crosslingual settings (e.g. ["en", "zh", "en-zh", "zh-en"]).

python BCEmbedding/tools/eval_mteb/eval_reranker_mteb.py --model_name_or_path maidalun1020/bce-reranker-base_v1

The evaluation tasks contain 12 datasets of "Reranking".

We provide a one-click script to summarize evaluation results of embedding and reranker models as Embedding Models Evaluation Summary and Reranker Models Evaluation Summary.

python BCEmbedding/evaluation/mteb/summarize_eval_results.py --results_dir {your_embedding_results_dir | your_reranker_results_dir}

LlamaIndex is a famous data framework for LLM-based applications, particularly in RAG. Recently, a LlamaIndex Blog has evaluated the popular embedding and reranker models in RAG pipeline and attracts great attention. Now, we follow its pipeline to evaluate our BCEmbedding.

First, install LlamaIndex, and upgrade transformers to 4.36.0:

pip install transformers==4.36.0

pip install llama-index==0.9.22

Export your "openai" and "cohere" app keys, and openai base url (e.g. "https://api.openai.com/v1") to env:

export OPENAI_BASE_URL={openai_base_url}  # https://api.openai.com/v1
export OPENAI_API_KEY={your_openai_api_key}
export COHERE_APPKEY={your_cohere_api_key}

• Hit Rate:

  Hit rate calculates the fraction of queries where the correct answer is found within the top-k retrieved documents. In simpler terms, it's about how often our system gets it right within the top few guesses. The larger, the better.

• Mean Reciprocal Rank (MRR):

  For each query, MRR evaluates the system's accuracy by looking at the rank of the highest-placed relevant document. Specifically, it's the average of the reciprocals of these ranks across all the queries. So, if the first relevant document is the top result, the reciprocal rank is 1; if it's second, the reciprocal rank is 1/2, and so on. The larger, the better.

In order to compare our BCEmbedding with other embedding and reranker models fairly, we provide a one-click script to reproduce results of the LlamaIndex Blog, including our BCEmbedding:

# There should be two GPUs available at least.
CUDA_VISIBLE_DEVICES=0,1 python BCEmbedding/tools/eval_rag/eval_llamaindex_reproduce.py

Then, summarize the evaluation results by:

python BCEmbedding/tools/eval_rag/summarize_eval_results.py --results_dir BCEmbedding/results/rag_reproduce_results

Results reproduced from the LlamaIndex Blog can be checked in Reproduced Summary of RAG Evaluation, with some obvious conclusions:

• In WithoutReranker setting, our bce-embedding-base_v1 outperforms all the other embedding models.
• With fixing the embedding model, our bce-reranker-base_v1 achieves the best performance.
• The combination of bce-embedding-base_v1 and bce-reranker-base_v1 is SOTA.

The evaluation of LlamaIndex Blog is monolingual, small amount of data, and specific domain (just including "llama2" paper). In order to evaluate the broad domain adaptability, bilingual and crosslingual capability, we follow the blog to build a multiple domains evaluation dataset (includding "Computer Science", "Physics", "Biology", "Economics", "Math", and "Quantitative Finance". Details), named CrosslingualMultiDomainsDataset:

• To prevent test data leakage, English eval data is selected from the latest English articles in various fields on ArXiv, up to date December 30, 2023. Chinese eval data is selected from high-quality, as recent as possible, Chinese articles in the corresponding fields on Semantic Scholar.
• Use OpenAI gpt-4-1106-preview to produce eval data for high quality.

First, run following cmd to evaluate the most popular and powerful embedding and reranker models:

# There should be two GPUs available at least.
CUDA_VISIBLE_DEVICES=0,1 python BCEmbedding/tools/eval_rag/eval_llamaindex_multiple_domains.py

Then, run the following script to summarize the evaluation results:

python BCEmbedding/tools/eval_rag/summarize_eval_results.py --results_dir BCEmbedding/results/rag_results

The summary of multiple domains evaluations can be seen in Multiple Domains Scenarios.

NOTE:

• Our bce-embedding-base_v1 outperforms other open-source embedding models with comparable model sizes.
• 114 datasets including 119 eval results (some dataset contains multiple languages) of "Retrieval", "STS", "PairClassification", "Classification", "Reranking" and "Clustering" in ["en", "zh", "en-zh", "zh-en"] setting, including MTEB and CMTEB.
• The crosslingual evaluation datasets we released belong to Retrieval task.
• More evaluation details should be checked in Embedding Models Evaluations.

NOTE:

• Our bce-reranker-base_v1 outperforms other open-source reranker models.
• 12 datasets of "Reranking" in ["en", "zh", "en-zh", "zh-en"] setting.
• More evaluation details should be checked in Reranker Models Evaluations.

NOTE:

• Data Quality:
  • To prevent test data leakage, English eval data is selected from the latest English articles in various fields on ArXiv, up to date December 30, 2023. Chinese eval data is selected from high-quality, as recent as possible, Chinese articles in the corresponding fields on Semantic Scholar.
  • Use OpenAI gpt-4-1106-preview to produce eval data for high quality.
• Evaluated in ["en", "zh", "en-zh", "zh-en"] setting. If you are interested in monolingual setting, please check in Chinese RAG evaluations with ["zh"] setting, and English RAG evaluations with ["en"] setting.
• Consistent with our Reproduced Results of LlamaIndex Blog.
• In WithoutReranker setting, our bce-embedding-base_v1 outperforms all the other embedding models.
• With fixing the embedding model, our bce-reranker-base_v1 achieves the best performance.
• The combination of bce-embedding-base_v1 and bce-reranker-base_v1 is SOTA.

For users who prefer a hassle-free experience without the need to download and configure the model on their own systems, BCEmbedding is readily accessible through Youdao's API. This option offers a streamlined and efficient way to integrate BCEmbedding into your projects, bypassing the complexities of manual setup and maintenance. Detailed instructions and comprehensive API documentation are available at Youdao BCEmbedding API. Here, you'll find all the necessary guidance to easily implement BCEmbedding across a variety of use cases, ensuring a smooth and effective integration for optimal results.

Welcome to scan the QR code below and join the WeChat group.

If you use BCEmbedding in your research or project, please feel free to cite and star it:

@misc{youdao_bcembedding_2023,
    title={BCEmbedding: Bilingual and Crosslingual Embedding for RAG},
    author={NetEase Youdao, Inc.},
    year={2023},
    howpublished={\url{https://github.com/netease-youdao/BCEmbedding}}
}

BCEmbedding is licensed under Apache 2.0 License

Netease Youdao - QAnything

FlagEmbedding

MTEB

C_MTEB

LLama Index | LlamaIndex Blog

HuixiangDou