ColBERT

Late interaction is the idea ColBERT introduced: encode queries and passages into token-level embedding matrices and match them with fast MaxSim operations at search time, retaining BERT-quality relevance while searching large collections in tens of milliseconds. The Stanford FutureData repository implements ColBERTv2, which compresses embeddings aggressively and pairs with the PLAID engine for efficient indexing and serving; a checkpoint pretrained on MS MARCO Passage Ranking is provided, so most users index their own corpus without any training. Installation is pip install colbert-ai with torch and faiss-gpu extras, or the provided conda environments for more reliable dependency resolution; the README states a GPU is required for training and indexing, while a CPU-only conda environment covers inference and a lightweight server.py ships for deployment. Five peer-reviewed papers from SIGIR 2020 through EMNLP 2023 document the line of work, and the MIT license permits commercial use. With 3.9k stars, it is the foundation of the late-interaction paradigm now spreading through rerankers and multimodal retrievers.