Open-Retrievals は、PyTorch と Transformers をベースとした、情報検索と LLM 検索拡張生成を指向した、SOTA テキスト埋め込みを取得する使いやすい Python フレームワークです。

AutoModelForEmbedding はベクトル化、検索、リランクの分野を統一します
対照学習エンベッディング, LLM エンベッディング
高速 RAG デモ

Exp	Model	Size	Original	Finetuned
embed pairwise finetune	bge-base-zh-v1.5	-	0.657	0.703
embed LLM finetune (LoRA)	e5-mistral-7b-instruct	-	0.651	0.699
rerank cross encoder	bge-reranker-base	-	0.666	0.706
rerank colbert	bge-m3	-	0.657	0.695
rerank LLM (LoRA)	bge-reranker-v2-gemma	-	0.637	0.706

インストール

必須条件

pip install transformers
pip install faiss-cpu  # 必要な場合
pip install peft  # 必要な場合

With pip

pip install open-retrievals

クイックスタート

事前訓練されたウェイトを使用

from retrievals import AutoModelForEmbedding

sentences = [
    'query: how much protein should a female eat',
    'query: summit define',
    "passage: As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "passage: Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]
model_name_or_path = 'intfloat/e5-base-v2'
model = AutoModelForEmbedding.from_pretrained(model_name_or_path, pooling_method="mean")
embeddings = model.encode(sentences, normalize_embeddings=True)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print(scores.tolist())

インデックスの構築と検索

from retrievals import AutoModelForEmbedding, AutoModelForRetrieval

sentences = ['A dog is chasing car.', 'A man is playing a guitar.']
model_name_or_path = "sentence-transformers/all-MiniLM-L6-v2"
index_path = './database/faiss/faiss.index'
model = AutoModelForEmbedding.from_pretrained(model_name_or_path, pooling_method='mean')
model.build_index(sentences, index_path=index_path)

query_embed = model.encode("He plays guitar.")
matcher = AutoModelForRetrieval()
dists, indices = matcher.search(query_embed, index_path=index_path)
print(indices)

リランク

from retrievals import AutoModelForRanking

model_name_or_path: str = "BAAI/bge-reranker-base"
rerank_model = AutoModelForRanking.from_pretrained(model_name_or_path)
scores_list = rerank_model.compute_score(["In 1974, I won the championship in Southeast Asia in my first kickboxing match", "In 1982, I defeated the heavy hitter Ryu Long."])
print(scores_list)

LangChain を使用した RAG

pip install langchain
pip install langchain_community
pip install chromadb

サーバー

from retrievals.tools.langchain import LangchainEmbedding, LangchainReranker, LangchainLLM
from retrievals import AutoModelForRanking
from langchain.retrievers import ContextualCompressionRetriever
from langchain.prompts.prompt import PromptTemplate
from langchain.chains import RetrievalQA
from langchain_chroma import Chroma as Vectorstore

persist_directory = './database/faiss.index'
embed_model_name_or_path = "sentence-transformers/all-MiniLM-L6-v2"
rerank_model_name_or_path = "BAAI/bge-reranker-base"
llm_model_name_or_path = "microsoft/Phi-3-mini-128k-instruct"

embeddings = LangchainEmbedding(model_name=embed_model_name_or_path)
vectordb = Vectorstore(
    collection_name="example_collection",
    persist_directory=persist_directory,
    embedding_function=embeddings,
)
retrieval_args = {"search_type" :"similarity", "score_threshold": 0.15, "k": 10}
retriever = vectordb.as_retriever(**retrieval_args)

ranker = AutoModelForRanking.from_pretrained(rerank_model_name_or_path)
reranker = LangchainReranker(model=ranker, top_n=3)
compression_retriever = ContextualCompressionRetriever(
    base_compressor=reranker, base_retriever=retriever
)

llm = LangchainLLM(model_name_or_path=llm_model_name_or_path)

RESPONSE_TEMPLATE = """[INST]
<>
You are a helpful AI assistant. Use the following pieces of context to answer the user's question.<>
Anything between the following `context` html blocks is retrieved from a knowledge base.

    {context}

REMEMBER:
- If you don't know the answer, just say that you don't know, don't try to make up an answer.
- Let's take a deep breath and think step-by-step.

Question: {question}[/INST]
Helpful Answer:
"""

PROMPT = PromptTemplate(template=RESPONSE_TEMPLATE, input_variables=["context", "question"])

qa_chain = RetrievalQA.from_chain_type(
    llm,
    chain_type='stuff',
    retriever=compression_retriever,
    chain_type_kwargs={
        "verbose": True,
        "prompt": PROMPT,
    }
)

user_query = 'Introduce this'
response = qa_chain({"query": user_query})
print(response)

コントラスト学習による transformers のウェイトのファインチューニング

import torch.nn as nn
from datasets import load_dataset
from transformers import AutoTokenizer, AdamW, get_linear_schedule_with_warmup, TrainingArguments
from retrievals import AutoModelForEmbedding, RetrievalTrainer, PairCollator, TripletCollator
from retrievals.losses import ArcFaceAdaptiveMarginLoss, InfoNCE, SimCSE, TripletLoss

model_name_or_path: str = "sentence-transformers/paraphrase-multilingual-mpnet-base-v2"
batch_size: int = 128
epochs: int = 3

train_dataset = load_dataset('shibing624/nli_zh', 'STS-B')['train']
train_dataset = train_dataset.rename_columns({'sentence1': 'query', 'sentence2': 'document'})
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=False)
model = AutoModelForEmbedding.from_pretrained(model_name_or_path, pooling_method="mean")
model = model.set_train_type('pairwise')

optimizer = AdamW(model.parameters(), lr=5e-5)
num_train_steps = int(len(train_dataset) / batch_size * epochs)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0.05 * num_train_steps, num_training_steps=num_train_steps)

training_arguments = TrainingArguments(
    output_dir='./checkpoints',
    num_train_epochs=epochs,
    per_device_train_batch_size=batch_size,
    remove_unused_columns=False,
)
trainer = RetrievalTrainer(
    model=model,
    args=training_arguments,
    train_dataset=train_dataset,
    data_collator=PairCollator(tokenizer, query_max_length=128, document_max_length=128),
    loss_fn=InfoNCE(nn.CrossEntropyLoss(label_smoothing=0.05)),
)
trainer.optimizer = optimizer
trainer.scheduler = scheduler
trainer.train()

対照学習による埋め込み LLM のファインチューニング

from retrievals import AutoModelForEmbedding

model = AutoModelForEmbedding.from_pretrained(
    "mistralai/Mistral-7B-v0.1",
    pooling_method='last',
    query_instruction=f'Instruct: Retrieve semantically similar text\nQuery: '
)

リランク

from transformers import AutoTokenizer, TrainingArguments, get_cosine_schedule_with_warmup, AdamW
from retrievals import RerankCollator, AutoModelForRanking, RerankTrainer, RerankTrainDataset

model_name_or_path: str = "BAAI/bge-reranker-base"
max_length: int = 128
learning_rate: float = 3e-5
batch_size: int = 4
epochs: int = 3
output_dir: str = "./checkpoints"

train_dataset = RerankTrainDataset(
    "C-MTEB/T2Reranking", positive_key="positive", negative_key="negative", dataset_split='dev'
)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=False)
model = AutoModelForRanking.from_pretrained(model_name_or_path)
optimizer = AdamW(model.parameters(), lr=learning_rate)
num_train_steps = int(len(train_dataset) / batch_size * epochs)
scheduler = get_cosine_schedule_with_warmup(
    optimizer,
    num_warmup_steps=0.05 * num_train_steps,
    num_training_steps=num_train_steps,
)

training_args = TrainingArguments(
    learning_rate=learning_rate,
    per_device_train_batch_size=batch_size,
    num_train_epochs=epochs,
    output_dir=output_dir,
    remove_unused_columns=False,
    logging_steps=100,
    report_to="none",
)
trainer = RerankTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    data_collator=RerankCollator(tokenizer, max_length=max_length),
)
trainer.optimizer = optimizer
trainer.scheduler = scheduler
trainer.train()

コサイン類似度/KNN による検索

from retrievals import AutoModelForEmbedding, AutoModelForRetrieval

query_texts = ['A dog is chasing car.']
document_texts = ['A man is playing a guitar.', 'A bee is flying low']
model_name_or_path = "sentence-transformers/all-MiniLM-L6-v2"
model = AutoModelForEmbedding.from_pretrained(model_name_or_path)
query_embeddings = model.encode(query_texts, convert_to_tensor=True)
document_embeddings = model.encode(document_texts, convert_to_tensor=True)

matcher = AutoModelForRetrieval(method='cosine')
dists, indices = matcher.search(query_embeddings, document_embeddings, top_k=1)

参考資料と謝辞

UKPLab/sentence-transformers
luyug/Dense
FlagOpen/FlagEmbedding

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README_ja-JP.md

README_ja-JP.md

インストール

クイックスタート

参考資料と謝辞

Files

README_ja-JP.md

Latest commit

History

README_ja-JP.md

File metadata and controls

インストール

クイックスタート

参考資料と謝辞