Neural-Cherche is a Python library for fine-tuning and deploying neural search models like Splade, ColBERT, and SparseEmbed. It targets researchers and developers needing to adapt state-of-the-art retrieval systems to specific datasets for improved performance in offline and online applications. The library simplifies the process of training, inference, and embedding management.

How It Works

Neural-Cherche facilitates fine-tuning using a triplet loss approach (anchor, positive, negative) on datasets formatted as tuples. It supports ColBERT fine-tuning from any Sentence Transformer checkpoint and Splade/SparseEmbed from MLM pre-trained models. The library also provides efficient inference classes for both retrieval and ranking stages, enabling users to build hybrid search systems. It allows saving computed embeddings to avoid redundant calculations.

Quick Start & Requirements

• Install: pip install neural-cherche or pip install "neural-cherche[eval]" for evaluation during training.
• Prerequisites: Python 3.x, PyTorch. GPU or MPS device recommended for training.
• Documentation: https://neural-cherche.readthedocs.io/en/latest/

Highlighted Details

• Supports CPU, GPU, and MPS devices.
• Provides pre-trained checkpoints for ColBERT and SparseEmbed on MS-MARCO.
• Includes implementations for BM25, TFIDF, SparseEmbed, SPLADE, and ColBERT.
• Offers a hybrid retrieval pipeline combining BM25 with ColBERT ranking for state-of-the-art results on benchmarks like SciFact.

Maintenance & Community

• Contributors: Benjamin Clavié, Arthur Satouf.
• References key papers for SPLADE, SparseEmbed, and ColBERT.

Licensing & Compatibility

• Library License: MIT.
• Model Licenses: Splade model is non-commercial only. SparseEmbed and ColBERT are fully open-source, including for commercial use.

Limitations & Caveats

The Splade model is restricted to non-commercial use, which may impact its applicability in certain enterprise environments. Fine-tuning Splade and SparseEmbed requires MLM pre-trained models, adding a dependency on specific model architectures.