open-retrievals

Documentation | Tutorials | 中文

Open-Retrievals is an easy-to-use python framework getting SOTA text embeddings, oriented to information retrieval and LLM retrieval augmented generation, based on PyTorch and Transformers.

• Contrastive learning enhanced embeddings
• LLM embeddings

Installation

Prerequisites

pip install transformers
pip install faiss-cpu
pip install peft

With pip

pip install open-retrievals

Quick-start

from retrievals import AutoModelForEmbedding, AutoModelForRetrieval

# Example list of documents
documents = [
    "Open-retrievals is a text embedding libraries",
    "I can use it simply with a SOTA RAG application.",
]

# This will trigger the model download and initialization
model_name_or_path = "sentence-transformers/all-MiniLM-L6-v2"
model = AutoModelForEmbedding(model_name_or_path)

embeddings = model.encode(documents)
len(embeddings) # Vector of 384 dimensions

Usage

Build Index and Retrieval

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"
model = AutoModelForEmbedding(model_name_or_path)
model.build_index(sentences)

matcher = AutoModelForRetrieval()
results = matcher.faiss_search("He plays guitar.")

Rerank

from transformers import AutoTokenizer
from retrievals import RerankCollator, RerankModel, RerankTrainer, RerankDataset

train_dataset = RerankDataset(args=data_args)
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, use_fast=False)

model = RerankModel(
    model_args.model_name_or_path,
    pooling_method="mean"
)
optimizer = get_optimizer(model, lr=5e-5, weight_decay=1e-3)

lr_scheduler = get_scheduler(optimizer, num_train_steps=int(len(train_dataset) / 2 * 1))

trainer = RerankTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    data_collator=RerankCollator(tokenizer, max_length=data_args.query_max_len),
)
trainer.optimizer = optimizer
trainer.scheduler = lr_scheduler
trainer.train()

RAG with LangChain

• Prerequisites

pip install langchain

• Server

from retrievals.tools.langchain import LangchainEmbedding, LangchainReranker
from langchain.retrievers import ContextualCompressionRetriever
from langchain_community.vectorstores import Chroma as Vectorstore


class DenseRetrieval:
    def __init__(self, persist_directory):
        embeddings = LangchainEmbedding(model_name="BAAI/bge-large-zh-v1.5")
        vectordb = Vectorstore(
            persist_directory=persist_directory,
            embedding_function=embeddings,
        )
        retrieval_args = {"search_type" :"similarity", "score_threshold": 0.15, "k": 30}
        self.retriever = vectordb.as_retriever(retrieval_args)

        reranker_args = {
            "model": "../../inputs/bce-reranker-base_v1",
            "top_n": 7,
            "device": "cuda",
            "use_fp16": True,
        }
        self.reranker = LangchainReranker(**reranker_args)
        self.compression_retriever = ContextualCompressionRetriever(
            base_compressor=self.reranker, base_retriever=self.retriever
        )

    def query(
        self,
        question: str
    ):
        docs = self.compression_retriever.get_relevant_documents(question)
        return docs

Use Pretrained sentence embedding

from retrievals import AutoModelForEmbedding

sentences = ["Hello world", "How are you?"]
model_name_or_path = "sentence-transformers/all-MiniLM-L6-v2"
model = AutoModelForEmbedding(model_name_or_path, pooling_method="mean", normalize_embeddings=True)
sentence_embeddings = model.encode(sentences, convert_to_tensor=True)
print(sentence_embeddings)

Finetune transformers by contrastive learning

from transformers import AutoTokenizer
from retrievals import AutoModelForEmbedding, AutoModelForRetrieval, RetrievalTrainer, PairCollator, TripletCollator
from retrievals.losses import ArcFaceAdaptiveMarginLoss, InfoNCE, SimCSE, TripletLoss
from retrievals.data import  RetrievalDataset, RerankDataset


train_dataset = RetrievalDataset(args=data_args)
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, use_fast=False)

model = AutoModelForEmbedding(
    model_args.model_name_or_path,
    pooling_method="cls"
)
optimizer = get_optimizer(model, lr=5e-5, weight_decay=1e-3)

lr_scheduler = get_scheduler(optimizer, num_train_steps=int(len(train_dataset) / 2 * 1))

trainer = RetrievalTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    data_collator=TripletCollator(tokenizer, max_length=data_args.query_max_len),
    loss_fn=TripletLoss(),
)
trainer.optimizer = optimizer
trainer.scheduler = lr_scheduler
trainer.train()

Finetune LLM for embedding by Contrastive learning

from retrievals import AutoModelForEmbedding

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

Search by Cosine similarity/KNN

from retrievals import AutoModelForEmbedding, AutoModelForRetrieval

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

matcher = AutoModelForRetrieval(method='cosine')
dists, indices = matcher.similarity_search(query_embeddings, passage_embeddings, top_k=1)

Reference & Acknowledge

• sentence-transformers
• FlagEmbedding
• uniem
• BCEmbedding