itinai content

←back to Blog

Liquid AI Releases LFM2-ColBERT-350M: A New Small Model that brings Late Interaction Retrieval to Multilingual and Cross-Lingual RAG

Liquid AI Releases LFM2-ColBERT-350M: A New Small Model that Brings Late Interaction Retrieval to Multilingual and Cross-Lingual RAG

Target Audience Analysis

The target audience for the LFM2-ColBERT-350M model primarily includes data scientists, AI researchers, and business managers in tech companies focused on natural language processing (NLP) and information retrieval. These professionals are often tasked with improving search functionalities across various languages and enhancing user experience in multilingual environments.

Pain Points: The audience faces challenges in efficiently retrieving relevant information across different languages, managing large datasets, and ensuring fast inference times without sacrificing accuracy.

Goals: They aim to implement advanced retrieval systems that support multilingual capabilities, improve search accuracy, and optimize performance in cross-lingual contexts.

Interests: This audience is interested in the latest advancements in AI models, particularly those that enhance retrieval-augmented generation (RAG) systems, and they seek practical applications of these technologies in business settings.

Communication Preferences: They prefer technical content that is concise, data-driven, and includes clear specifications and use cases. Visual aids such as diagrams and charts are also beneficial for understanding complex concepts.

Overview of LFM2-ColBERT-350M

Liquid AI has introduced the LFM2-ColBERT-350M, a compact late interaction retriever designed for multilingual and cross-lingual search. This model allows documents to be indexed in one language while enabling queries in multiple languages, achieving high accuracy in retrieval. The inference speed of LFM2-ColBERT-350M is comparable to models that are 2.3 times smaller, a performance attributed to its LFM2 backbone. A demo is available on Hugging Face, along with a detailed model card for integration into retrieval-augmented generation systems.

Understanding Late Interaction

Late interaction retrieval combines the speed of bi-encoders with the accuracy of cross-encoders. In this approach, queries and documents are encoded separately at the token level. At query time, token vectors are compared using operations such as MaxSim, which preserves fine-grained token interactions without the full computational cost of joint cross attention. This method allows for pre-computation of document embeddings, thereby enhancing precision during the ranking phase. It can function as both a first-stage retriever and a ranker in a single pass.

Model Specifications

  • Total Parameters: 350 million
  • Layers: 25
  • Convolution Blocks: 18
  • Attention Blocks: 6
  • Dense Layer: 1
  • Context Length: 32,000 tokens
  • Vocabulary Size: 65,536
  • Similarity Function: MaxSim
  • Output Dimensionality: 128
  • Training Precision: BF16
  • License: LFM Open License v1.0

Supported Languages

The LFM2-ColBERT-350M model supports eight languages: English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish. Additionally, evaluations include Italian and Portuguese, expanding the model’s capability for cross-lingual comparisons.

Evaluation Setup and Key Results

Liquid AI has extended the NanoBEIR benchmark to include Japanese and Korean, ensuring reproducibility of results. The LFM2-ColBERT-350M demonstrates superior multilingual capabilities compared to the baseline late interaction model, GTE-ModernColBERT-v1, which has 150 million parameters. Significant performance improvements were noted in German, Arabic, Korean, and Japanese, while maintaining English performance.

Key Takeaways

  • Token-level scoring with MaxSim preserves fine-grained interactions while maintaining separate encoders, allowing for efficient precomputation of document embeddings.
  • Documents can be indexed in one language and retrieved in multiple languages, supporting eight languages with evaluations across nine for cross-lingual pairs.
  • On the NanoBEIR multilingual extension, LFM2-ColBERT-350M outperforms the previous late interaction baseline and retains strong performance in English.
  • Inference speed is comparable to models 2.3 times smaller, thanks to the LFM2 backbone.

Conclusion

The LFM2-ColBERT-350M model by Liquid AI applies late interaction retrieval techniques with MaxSim, encoding queries and documents separately to score token vectors efficiently. This model is poised for deployment in multilingual retrieval-augmented generation trials, offering a scalable solution for businesses seeking to enhance their search capabilities.

For more information, check out the Model Weights, Demo, and Technical Details. You can also explore our GitHub Page for Tutorials, Codes, and Notebooks. Follow us on Twitter, join our 100k+ ML SubReddit, and subscribe to our Newsletter. If you’re on Telegram, you can join us there as well.