ColBERT
Also called Contextualized Late Interaction over BERT
A retrieval model that keeps a vector per word and matches them individually for sharper results.
Think of it like
Comparing two essays sentence by sentence instead of judging each by a single one-line summary.
Example
ColBERT retrieves passages by matching each query token to its best-matching document token, catching nuances a single-vector search misses.
How it actually works
ColBERT embeds every token of query and document separately, then scores a pair by summing each query token’s best match against the document’s tokens — "late interaction". This retains fine-grained signal that single-vector encoders lose, boosting quality, especially on nuanced queries. The price is much larger indexes and heavier computation, which is why it is powerful but not always the default.
For product teams
A quality upgrade for demanding search, weighed against bigger storage and infra cost.
For engineers
Token-level embeddings with MaxSim late interaction; higher accuracy at the cost of large multi-vector indexes.
Related
- Late Interaction — The flagship late interaction method.
- Multi-Vector Retrieval — A form of multi-vector retrieval.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome