RAGatouille (ColBERT made easy) ColBERT represents each document and query as a set of contextual token vectors and scores with late interaction (MaxSim). Recall is typically higher than single-vector dense retrieval for long documents and entity-heavy queries. RAGatouille is the pragmatic wrapper: one function to index, one to search, one to train. Installation Load a Pretrained Model Other checkpoints worth trying: (multilingual, 8192-token context), (smaller + fast), (2024, strong baseline). Index a Corpus Indexing a million chunks takes 1–4 hours on an A10; disk footprint is 10–20 % of ra…