late interaction models

**Late interaction models** is the **retrieval model family that delays document-query interaction to token-level matching after independent encoding** - it aims to combine high retrieval quality with scalable indexing. **What Is Late interaction models?** - **Definition**: Architecture storing multiple token representations per document and computing relevance at query time via token-level similarity aggregation. - **Interaction Pattern**: Stronger than single-vector bi-encoder scoring, lighter than full cross-encoder encoding. - **Typical Mechanism**: MaxSim-style matching between query tokens and document token embeddings. - **System Tradeoff**: Higher storage and scoring cost than bi-encoders, lower than exhaustive cross-encoder ranking. **Why Late interaction models Matters** - **Quality Improvement**: Captures finer semantic alignment and term-specific relevance. - **Retrieval Robustness**: Handles nuanced phrasing and partial lexical overlap better than single-vector methods. - **Scalable Precision**: Offers strong ranking quality without full pairwise transformer passes. - **RAG Benefit**: Better candidate quality improves grounding and reduces hallucination risk. - **Research Momentum**: Important bridge architecture in modern neural IR evolution. **How It Is Used in Practice** - **Index Design**: Store compressed token embeddings with efficient ANN-compatible structures. - **Scoring Optimization**: Tune token interaction aggregation for latency and quality balance. - **Pipeline Placement**: Use as high-quality first-stage retriever or pre-rerank layer. Late interaction models is **a powerful retrieval paradigm between bi-encoder speed and cross-encoder accuracy** - token-level scoring delivers meaningful relevance gains for complex query-document matching.