Glossary term
Glossary term
Memory and Retrieval
Late-interaction retrieval model that computes token-level document embeddings and scores relevance via maximum similarity matching.
RAGatouille (open-source) wraps ColBERTv2 for easy integration into LangChain and LlamaIndex RAG pipelines - used by academic researchers and enterprise teams needing high-precision retrieval over specialised corpora.
Stanford's DSP (Demonstrate-Search-Predict) framework uses ColBERT for multi-hop retrieval across Wikipedia - retrieving supporting passages for complex questions that require reasoning over 2-3 documents.
ColBERT-v2 achieves state-of-the-art on BEIR benchmark across 18 retrieval tasks - used by legal-tech companies for precise statute and case-law retrieval where keyword and dense methods both produce false positives.