# ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction

**Authors:** Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, Matei Zaharia  
**arXiv:** [2112.01488](https://arxiv.org/abs/2112.01488)  
**Date:** 2021-12-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.01488v1/

## Key points

- Neural Information Retrieval (IR) has revolutionized search and other knowledge-intensive language tasks.
- Late interaction models produce multi-vector representations at the granularity of each token, decomposing relevance modeling into scalable token-level computations.
- ColBERTv2 has been introduced as a retriever that combines denoised supervision and residual compression to simultaneously improve quality and reduce the space footprint by 5--8$\times$.
- ColBERTv2 has achieved state-of-the-art retrieval quality both within and outside its training domain on a wide array of benchmarks.
- The new LoTTE benchmark is a resource for out-of-domain evaluation that focuses on queries with practical intent over long-tail topics.
- Several neural IR approaches leverage multi-vector representations such as Poly-encoders, PreTTR, MORES, COIL, SPLADE, and SPLADEv2.
- Vector compression for neural IR has also gained recent interest in compressing representations for IR.
- ColBERTv2 uses distillation from a cross-encoder and hard negative mining to boost quality beyond any existing method before applying residual compression to reduce the space footprint while preserving quality.

## Summary

Neural Information Retrieval (IR) has revolutionized search and other knowledge-intensive language tasks. While many neural IR methods encode queries and documents into single-vector representations, late interaction models produce multi-vector representations at the granularity of each token, decomposing relevance modeling into scalable token-level computations. To address the challenge of space footprint inflation by an order of magnitude posed by this approach, ColBERTv2 has been introduced as a retriever that combines denoised supervision and residual compression to simultaneously improve quality and reduce the space footprint by 5--8$\times$. ColBERTv2 has achieved state-of-the-art retrieval quality both within and outside its training domain on a wide array of benchmarks. The new LoTTE benchmark is a resource for out-of-domain evaluation that focuses on queries with practical intent over long-tail topics. Several neural IR approaches leverage multi-vector representations such as Poly-encoders, PreTTR, MORES, COIL, SPLADE, and SPLADEv2. Vector compression for neural IR has also gained recent interest in compressing representations for IR. ColBERTv2 uses distillation from a cross-encoder and hard negative mining to boost quality beyond any existing method before applying residual compression to reduce the space footprint while preserving quality.