A model and package for German ColBERT

TL;DR

This paper presents a German adaptation of ColBERT and an accompanying software package to support embedding, indexing, training, and negative sampling for retrieval-augmented generation. It relies on the MaxSim token-level late-interaction score, defined as $\operatorname{MaxSim}(q,d) = \sum_{i=1}^m \max_{j} S(q_i, d_j)$, with a German BERT backbone trained on MS MARCO Passage Ranking translated to German using random negatives to maximize recall. Key contributions include releasing colbertkit for end-to-end ColBERT workflows and showing recall and NDCG gains over BM25 on German datasets. The work enables scalable, interpretable, token-level retrieval for German RAG and related search tasks.

Abstract

In this work, we introduce a German version for ColBERT, a late interaction multi-dense vector retrieval method, with a focus on RAG applications. We also present the main features of our package for ColBERT models, supporting both retrieval and fine-tuning workflows.