Late Interaction
Definition
Late interaction calculates fine-grained similarities between query and document representations only late in the retrieval process.
In ColBERT, token representations are precomputed separately; for search, query and document tokens are compared via MaxSim-like operations. This combines detail with better precomputability.
Simply explained
How to prepare two texts individually first and only compare their individual terms in detail during search.
Why it matters in practice
Relevant for high-quality semantic search and reranking.
Example from business
A document search uses late interaction when a single vector per document loses too much detail.
Limits and pitfalls
- Assuming higher quality without memory/latency costs; equate ColBERT with cross-encoder.
Origin & history
Khattab and Zaharia introduced ColBERT in 2020 as efficient BERT-based retrieval with late interaction.
Use in mid-sized companies
- Possible application in medium-sized businesses: Benchmark for demanding enterprise search against single-vector and cross-encoder approaches.
Frequently asked questions
Is late interaction a cross-encoder?
No. What is the advantage? More fine-grained matching signals with precomputable document representations.
Sources
- ColBERT · Khattab & Zaharia (2020), ColBERT
Sources checked on 10.10.2026.
Editorially reviewed on: