ColBERT is a search model that compares queries and documents word by word using many small embeddings, rather than squeezing each into one vector.
In plain words
Many search systems turn a whole passage into a single embedding. ColBERT keeps one embedding per token. It encodes documents ahead of time, then at search time encodes the query and, for each query token, finds its best match in the passage, then combines those best-match scores. This late interaction keeps fine detail while still letting documents be indexed in advance, so it is more accurate than single-vector search and still fast.
A quick example
For the query "side effects of ibuprofen", ColBERT lets each query word find its strongest match in a passage, so a passage covering both the drug and its side effects scores highest.
Where you'll meet it
- Retrieval for RAG systems that need high accuracy.
- Reranking or end-to-end passage search.
Related
Sources
- PaperColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT, Khattab and Zaharia (arXiv) · read 29 Sept 2026
- DocsColBERT, Stanford FutureData (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.