Beyond Single Vectors: Multi-Vector Search for Enhanced...
summary
Multi-vector late-interaction (ColBERT / ColPaLi) keeps token/patch vectors and uses max-sim token→token scoring, implemented in OpenSearch via nested vector fields and a scripting reranker. On an 8,000‑page PDF-as-image RAG demo, multi-vector retrieval outperforms single‑vector parsing by ~20 points while the Vidor leaderboard’s top 10 are multi‑vector. Tradeoffs: storage/memory rises ~25× and latency can spike to ~1s, mitigated by two‑phase retrieval (KNN oversample ~10× then token rerank) plus pooling/clustering and OpenSearch quantization up to 32×.
AI-written summary · about
chapters (10)
- 0:00Multi-vector search intro
- 8:48Late interaction models
- 13:27Late interaction evolution
- 19:10Explainability with multi-vectors
- 21:59Benchmarking relevance
- 24:27Production challenges
- 26:34Two-phase retrieval
- 32:02Live demo
- 38:48Implementation resources
- 39:31Q&A
related talks
01h 09mVector Search: Ask Me Anything!
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...
46mNatural Language & Search, Improve your Search w/ Vector capabilities
18mLate Interaction Retrieval in Practice: Implications for OpenSearch-Based Search Systems
59mContextual search with vector search: exploring your options with open source tools
19mVectorize Your Open Source Search Engine
33mArchitecting Scalable Multi-Modal Video Search