ColBERTv2

Overview

ColBERTv2 (Contextual Late Interactions BERT) is an advanced bi-encoder that represents text in multi-vector embeddings at the token level, introducing a novel late interaction mechanism for efficient and precise ranking.

How It Works

• Each token gets its own vector embedding
• Encodes passages into matrices of token-level embeddings
• Encodes queries into separate matrices
• Uses scalable MaxSim operators for similarity
• Late interaction compares vectors at final stage

Improvements Over ColBERT v1

• Denoised Supervision: Distillation from larger teacher models
• Hard Negative Mining: Improved training with challenging examples
• Reduced Memory: Addresses original's huge storage requirements
• Better Quality: Enhanced retrieval accuracy
• Improved Efficiency: Optimized for production deployment

Technical Architecture

• Multi-vector representation per document
• Token-level granularity
• Late interaction scoring
• MaxSim aggregation
• Efficient indexing strategies

Advantages

• Captures complex query semantics
• More accurate than single-vector bi-encoders
• Handles nuanced information needs
• Scalable to large collections
• Fine-grained matching capabilities

Database Support (2026)

• Weaviate 1.29+ (multi-vector embeddings)
• Qdrant with FastEmbed
• Vespa with native ColBERT embedder
• Milvus with Array of Structs
• VectorChord with MaxSim operators

Use Cases

• Precision-critical search
• Complex question answering
• Long-document retrieval
• Multi-hop reasoning
• Academic paper search

Performance

Significant advancement over traditional bi-encoder models, with multi-embedding representations clearly capturing complex queries for more accurate retrievals.

Pricing

Open-source model, self-hostable or via API providers