Description

Simon Willison delves into the workings of ColBERT, a retrieval model designed for scalable BERT-based search over extensive text collections. He explains how ColBERT differs from regular embedding models and how it provides more information than traditional embedding search by showing which words in the document are most relevant. Willison then proceeds to use RAGatouille, a library that makes working with ColBERT easier, to create an index of his blog's content. He also demonstrates querying the index and implementing a basic question-answering mechanism using LLM. The article further explores re-ranking queries without building an index first.