MaxSim Operator

Maximum Similarity late interaction function introduced by ColBERT for ranking. Calculates cosine similarity between query and document token embeddings, keeping maximum score per query token for highly effective long-document retrieval.

Library built on Sentence Transformers for flexible training, inference, and retrieval with state-of-the-art ColBERT models. Features FastPLAID index for efficient multi-vector late interaction retrieval with 10x storage compression and sub-200ms latency.

Embedding approach where documents/images are represented by multiple vectors (one per token/patch) rather than a single vector, enabling fine-grained semantic matching.

A retrieval paradigm where query and document encodings are kept separate until a late interaction stage, enabling more expressive and efficient similarity computations. Pioneered by ColBERT and extended by ColPali and ColQwen, this approach maintains fine-grained representations while enabling fast retrieval.

A two-stage retrieval process where initial candidates from vector search are reordered using more sophisticated models like cross-encoders. Reranking significantly improves result quality by applying computationally expensive models to a small set of candidates, commonly used in RAG systems and search applications.

Search technique combining BM25 lexical search and semantic vector search using Reciprocal Rank Fusion (RRF) to merge results, balancing precision of keyword matching with contextual understanding of neural embeddings.