neural-cherche 1.1.0

Sparse Embeddings for Neural Search.

Neural-Cherche

Neural Search

Neural-Cherche is a library designed to fine-tune neural search models such as Splade, ColBERT, and SparseEmbed on a specific dataset. Neural-Cherche also provide classes to run efficient inference on a fine-tuned retriever or ranker. Neural-Cherche aims to offer a straightforward and effective method for fine-tuning and utilizing neural search models in both offline and online settings. It also enables users to save all computed embeddings to prevent redundant computations.

Neural-Cherche is compatible with CPU, GPU and MPS devices. We can fine-tune ColBERT from any Sentence Transformer pre-trained checkpoint. Splade and SparseEmbed are more tricky to fine-tune and need a MLM pre-trained model.

Installation

We can install neural-cherche using:

pip install neural-cherche

If we plan to evaluate our model while training install:

pip install "neural-cherche[eval]"

Documentation

The complete documentation is available here.

Quick Start

Your training dataset must be made out of triples (anchor, positive, negative) where anchor is a query, positive is a document that is directly linked to the anchor and negative is a document that is not relevant for the anchor.

X = [
    ("anchor 1", "positive 1", "negative 1"),
    ("anchor 2", "positive 2", "negative 2"),
    ("anchor 3", "positive 3", "negative 3"),
]

And here is how to fine-tune ColBERT from a Sentence Transformer pre-trained checkpoint using neural-cherche:

import torch

from neural_cherche import models, utils, train

model = models.ColBERT(
    model_name_or_path="raphaelsty/neural-cherche-colbert",
    device="cuda" if torch.cuda.is_available() else "cpu" # or mps
)

optimizer = torch.optim.AdamW(model.parameters(), lr=3e-6)

X = [
    ("query", "positive document", "negative document"),
    ("query", "positive document", "negative document"),
    ("query", "positive document", "negative document"),
]

for step, (anchor, positive, negative) in enumerate(utils.iter(
        X,
        epochs=1, # number of epochs
        batch_size=8, # number of triples per batch
        shuffle=True
    )):

    loss = train.train_colbert(
        model=model,
        optimizer=optimizer,
        anchor=anchor,
        positive=positive,
        negative=negative,
        step=step,
        gradient_accumulation_steps=50,
    )

    
    if (step + 1) % 1000 == 0:
        # Save the model every 1000 steps
        model.save_pretrained("checkpoint")

Retrieval

Here is how to use the fine-tuned ColBERT model to retrieve documents:

from neural_cherche import models, retrieve
import torch

batch_size = 32

documents = [
    {"id": 0, "document": "Food"},
    {"id": 1, "document": "Sports"},
    {"id": 2, "document": "Cinema"},
]

model = models.ColBERT(
    model_name_or_path="raphaelsty/neural-cherche-colbert",
    device="cuda" if torch.cuda.is_available() else "cpu", # or mps
)

retriever = retrieve.ColBERT(
    key="id",
    on=["document"], # the field to search on, can be a list of fields
    model=model,
)

documents_embeddings = retriever.encode_documents(
    documents=documents,
    batch_size=batch_size,
)

retriever = retriever.add(
    documents_embeddings=documents_embeddings,
)

Now we can retrieve documents using the fine-tuned model:

queries_embeddings = retriever.encode_queries(
    queries=["Food", "Sports", "Cinema"], # list of queries
    batch_size=batch_size,
)

scores = retriever(
    queries_embeddings=queries_embeddings,
    batch_size=batch_size,
    k=10, # number of documents to retrieve
)

scores

[[{'id': 0, 'similarity': 22.825355529785156},
  {'id': 1, 'similarity': 11.201947212219238},
  {'id': 2, 'similarity': 10.748161315917969}],
 [{'id': 1, 'similarity': 23.21628189086914},
  {'id': 0, 'similarity': 9.9658203125},
  {'id': 2, 'similarity': 7.308732509613037}],
 [{'id': 1, 'similarity': 6.4031805992126465},
  {'id': 0, 'similarity': 5.601611137390137},
  {'id': 2, 'similarity': 5.599479675292969}]]

Please note that neural-cherche provide documentation to use ColBERT as a ranker which is much more efficient.

Neural-Cherche also provides a SparseEmbed, a SPLADE, a TFIDF retriever and a ColBERT ranker which can be used to re-order output of a retriever. For more information, please refer to the documentation.

Pre-trained Models

We provide pre-trained checkpoints specifically designed for neural-cherche: raphaelsty/neural-cherche-sparse-embed and raphaelsty/neural-cherche-colbert. Those checkpoints are fine-tuned on a subset of the MS-MARCO dataset and would benefit from being fine-tuned on your specific dataset. You can fine-tune ColBERT from any Sentence Transformer pre-trained checkpoint in order to fit your specific language. You should use a MLM based-checkpoint to fine-tune SparseEmbed.

		scifact dataset
model	HuggingFace Checkpoint	ndcg@10	hits@10	hits@1
TfIdf	-	0,61	0,85	0,47
SparseEmbed	raphaelsty/neural-cherche-sparse-embed	0,62	0,87	0,48
Sentence Transformer	sentence-transformers/all-mpnet-base-v2	0,66	0,89	0,53
ColBERT	raphaelsty/neural-cherche-colbert	0,70	0,92	0,58
TfIDF Retriever + ColBERT Ranker	raphaelsty/neural-cherche-colbert	0,71	0,94	0,59

Neural-Cherche Contributors

• Benjamin Clavié

References

• SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking authored by Thibault Formal, Benjamin Piwowarski, Stéphane Clinchant, SIGIR 2021.

• SPLADE v2: Sparse Lexical and Expansion Model for Information Retrieval authored by Thibault Formal, Carlos Lassance, Benjamin Piwowarski, Stéphane Clinchant, SIGIR 2022.

• SparseEmbed: Learning Sparse Lexical Representations with Contextual Embeddings for Retrieval authored by Weize Kong, Jeffrey M. Dudek, Cheng Li, Mingyang Zhang, and Mike Bendersky, SIGIR 2023.

• ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT authored by Omar Khattab, Matei Zaharia, SIGIR 2020.

License

This Python library is licensed under the MIT open-source license, and the splade model is licensed as non-commercial only by the authors. SparseEmbed and ColBERT are fully open-source including commercial usage.