Advanced Vector Search for LLM Applications
Chapter 1: Approximate Nearest Neighbor Algorithms
Revisiting Vector Embeddings and Search Fundamentals
Hierarchical Navigable Small Worlds (HNSW) Internals
Inverted File Index (IVF) Variations
Product Quantization (PQ) Mechanics
Other Graph-Based ANN Methods (e.g., NSG, Vamana)
Selecting the Right ANN Algorithm: Trade-offs
Hands-on Practical: Implementing and Tuning HNSW
Chapter 2: Optimizing Vector Search Performance and Efficiency
Quantization Techniques: Scalar vs. Product
Implementing Optimized Product Quantization (OPQ)
Binary Hashing and Locality Sensitive Hashing (LSH) Refresher
Advanced Filtering Strategies: Pre vs. Post Filtering
Indexing Metadata Efficiently alongside Vectors
Hardware Acceleration Considerations (CPU SIMD, GPU)
Memory Management and Caching Strategies
Practice: Applying Quantization and Filtering
Chapter 3: Hybrid Search Approaches
Limitations of Pure Vector Search
Integrating Keyword Search (BM25, TF-IDF)
Result Fusion and Ranking Strategies
Reciprocal Rank Fusion (RRF) and Other Fusion Algorithms
Graph-Based Augmentation for Vector Search
Multi-Modal Search Considerations
Hands-on Practical: Building a Hybrid Search Pipeline
Chapter 4: Scaling Vector Search for Production Systems
Distributed Vector Database Architectures
Sharding Strategies for Vector Indexes
Replication and High Availability
Load Balancing Search Queries
Monitoring Vector Search Performance Metrics
Index Updates and Maintenance in Production
Cost Optimization for Large-Scale Deployments
Practice: Configuring a Distributed Setup
Chapter 5: Advanced Tuning and Evaluation
Evaluation Metrics Revisited: Recall, Precision, Latency
Building Ground Truth Datasets for Evaluation
Parameter Sensitivity Analysis (HNSW, IVF)
A/B Testing Frameworks for Search Algorithms
Debugging Search Relevance Issues
Online vs. Offline Evaluation Techniques
Tuning for Specific Application Needs (RAG vs. Semantic Search)
Hands-on Practical: Comprehensive Performance Evaluation
Sharding Strategies for Vector Indexes
Was this section helpful?
- Milvus: A Purpose-Built Vector Data Management System, Gu, Xinyu and Song, Jun and Zheng, Xiaomeng and Li, Bowen and Zhang, Bin and Chen, Xiaohan and Su, Jincheng and Wu, Weiping and Zhang, Chao and Chen, Xiaolong and Jiang, Gang and Hu, Yu, 2021 Proceedings of the VLDB Endowment, Vol. 14 (VLDB Endowment) DOI: 10.14778/3476239.3476274 - This paper details the architecture of Milvus, an open-source vector database, which includes sharding strategies and distributed query processing for large-scale vector search. It provides insights into how a production vector search system handles scalability.
- Designing Data-Intensive Applications, Martin Kleppmann, 2017 (O'Reilly Media) - Chapter 6, 'Partitioning Data,' offers an explanation of various sharding (partitioning) strategies, including hash-based and range-based methods, load balancing, and rebalancing. These are general principles applicable to distributed vector indexes.