← AI Terminology
ColBERT
ColBERT is a late-interaction retrieval model that encodes queries and documents into multi-vector token-level embeddings and scores with efficient MaxSim interactions.
It sits between single-vector bi-encoders and full cross-encoders in the quality/cost tradeoff.
It sits between single-vector bi-encoders and full cross-encoders in the quality/cost tradeoff.
Why It Matters in AI
Single vectors crush nuance; cross-encoders are expensive at scale. ColBERT’s late interaction improves recall over bi-encoders while remaining indexable — a strong advanced RAG retriever when tuned well.
Key Points
| Aspect | Description |
|---|---|
| Cons | Larger indexes than single-vector |
| Idea | Keep token vectors; MaxSim between query and doc tokens |
| Pros | Strong effectiveness; scalable with PLAID indexes |
| Origin | Khattab & Zaharia (2020) |
| Related | Dense retrieval, rerankers |
| Ecosystem | RAGatouille, PLAID, ColBERTv2 |
Simple Analogy
Instead of one summary card per book, keep a card per paragraph and match the best paragraph bits to each word of the question — richer matching, still pre-indexable.
Common Usage Examples
- RAGatouille ColBERT indexing
- ColBERTv2 checkpoints
- Compare recall@k vs DPR bi-encoder
- PLAID for fast late interaction
Summary
In short: ColBERT retrieves with token-level late interaction — finer semantic match than single vectors, cheaper than full cross-encoding at index scale.