Instead of compressing each side into just one vector, an implementation can retain multiple contextualized vectors. A later comparison combines fine-grained matches into a document score.
Separately encoding documents allows their representations to be prepared before a query arrives. The method balances detailed matching against storage and computation costs; quality still needs testing on the intended collection.
ELI5
Imagine preparing detailed notes for each document, then comparing a new question with those notes.
For example, a search for a company's operating costs can compare several parts of the question with relevant details in a report.
Why is the interaction called late?
The query and document are encoded separately before their representations are compared.
Does late interaction guarantee accurate retrieval?
No. Results depend on training, document representations, the collection and the scoring method.
