TechPrufer
Get started
Milvus
Milvus logoMilvus · Feature trail

IVF Vector Index

1 post · Dec 2019 – Jun 2026

What this trail covers

Milvus 2.0 introduced the IVF vector index, a widely used algorithm for Approximate Nearest Neighbor (ANN) search. This post details its inner workings, including K-means clustering for partitioning vectors into buckets, vector assignment to inverted lists, and optional compression encoding (SQ8, PQ). It also outlines the search process: calculating distances to centroids, selecting nearest clusters based on `nprobe`, and performing exact or approximate comparisons within those clusters. Milvus 2.0 also introduced HNSWlib integration for ANN search, offering a graph-based approach with hierarchical layers and navigable small worlds for efficient similarity searches in high-dimensional spaces. HNSWlib provides features like sublinear search time, dynamic updates, and memory efficiency, making it suitable for medium-scale applications. However, for higher scalability requirements or enterprise-level features, purpose-built vector databases like Milvus are recommended.

All Milvus trails
  1. How to Run 25 Million Image Vectors on Under 1GB of Memory in MilvusThis post details how to achieve extremely low memory usage (under 1GB) for 25 million 1280-dimensional vectors by leveraging FLAT index with FP16 storage, mmap for raw vector data, and aggressive scalar filtering before vector comparison. It highlights that the effectiveness of FLAT depends on the query pattern reducing the search space significantly, and that advanced ANN indexes may not be beneficial when scalar filters are highly selective. The post also discusses the trade-offs of FP16 (potential recall impact) and mmap (disk usage, cold query latency), and advises testing real-world performance beyond sizing tool estimates.
  2. Understanding IVF Vector Index: How It Works and When to Choose It Over HNSWThis post provides a deep dive into the IVF (Inverted File) vector index algorithm, explaining its core concepts, building process (K-means clustering, vector assignment, optional compression), and search mechanism. It elaborates on the trade-offs associated with hyperparameters like `nlist` and `nprobe`, offering practical guidelines for their selection and tuning. The article also details common IVF variants (IVF_FLAT, IVF_PQ, IVF_SQ8) and their respective use cases, focusing on the balance between performance, accuracy, and memory usage. This content directly expands on the existing 'IVF Vector Index' feature thread by providing a comprehensive explanation of its internal mechanics and practical application.
  3. Understanding Hierarchical Navigable Small Worlds (HNSW) for Vector SearchThis post provides a high-level overview and conceptual explanation of the Hierarchical Navigable Small Worlds (HNSW) algorithm for vector similarity search. It details the construction and search processes of both Navigable Small Worlds (NSW) and HNSW, explaining the hierarchical structure and its benefits for course-to-fine search. The post contrasts HNSW with naive similarity search and highlights its suitability as a default index for medium-sized, in-memory vector datasets. It references external resources for detailed mathematical analysis and a pedagogical Python implementation.
  4. Getting Started with HNSWlibThis post provides a detailed explanation of the HNSW algorithm and its implementation in the HNSWlib library. It covers the hierarchical layers, navigable small worlds concept, skip list-like structure, search efficiency, and insertion/maintenance processes of HNSW. It also walks through setting up HNSWlib, preparing data, building an index with parameters like `space`, `dim`, `max_elements`, `ef_construction`, `M`, and `ef`, and performing nearest neighbor searches using `knn_query`. The post concludes by comparing HNSWlib's capabilities with purpose-built vector databases like Milvus, highlighting HNSWlib's suitability for prototyping and medium-sized datasets while recommending Milvus for higher scalability and enterprise features.
  5. Efficient Vector Similarity Search in Recommender Workflows Using Milvus with NVIDIA MerlinThis post demonstrates the integration of Milvus with NVIDIA Merlin for efficient vector similarity search in recommender system workflows. It details how Milvus, with GPU acceleration leveraging NVIDIA RAFT, complements Merlin in the item retrieval stage by providing fast ANN search over vector embeddings generated by Merlin Models. The post includes a benchmark study showing substantial speedups (37x to 91x) with GPU-accelerated Milvus compared to CPU-based approaches, and outlines the tech stack and architectural considerations for such integrations.
  7. What Powers Similarity Search in Milvus Vector Database?This post introduces Knowhere, Milvus's core vector execution engine, detailing its concept, architecture within Milvus, and advantages over Faiss. It highlights Knowhere's support for bitset, additional similarity metrics for binary vectors, AVX512 instruction sets, and automatic SIMD instruction selection. The post also outlines the code structure of Knowhere, including its base classes and how to add new indexes.
  8. Quickly Test and Deploy Vector Search Solutions with the Milvus 2.0 BootcampThis post introduces the Milvus 2.0 bootcamp, which provides updated guides and code examples for testing and deploying vector search solutions. It includes benchmark tests for 1M and 100M datasets, and example solutions for image, video, audio similarity search, recommendation systems, molecular search, and question answering systems. A detailed step-by-step guide for an image similarity search example is provided, covering server setup, collection creation with IVF_SQ8 index, model integration, data insertion, and similarity search. The post also highlights Milvus 2.0's cloud-native architecture, real-time query capabilities, bitset usage for data management, and improved data file cleanup mechanisms.
  9. Building a Wardrobe and Outfit Planning App with MilvusThis post details the implementation of an image search system for a fashion app (Stylepedia) using Milvus. It describes the system architecture, including offline (vectorization, indexing) and online (querying) components. Key technical details include the use of YOLOv5 for garment detection, EfficientNet for feature extraction (512-dimensional vectors), and Milvus with the HNSW index for vector similarity search. The implementation utilizes partitioning based on garment classification to narrow search scope and achieves millisecond search times on tens of millions of vectors. Post-processing steps like color and key label filtering are also described. The application of this system for user uploads, outfit suggestions, and fashion photograph recommendations is illustrated.
  10. The Journey to Optimizing Billion-scale Image Search (2/2)This post details the second-generation search-by-image system using a CNN + Milvus solution. It explains feature extraction using VGG16 with Keras + TensorFlow, including normalization, image loading from bytes, and black border processing. It then discusses Milvus as a vector search engine, covering CPU requirements (AVX2), capacity planning for billion-scale datasets, database design with collections and partitions, structured data mapping, index selection, and processing search results (ID + distance, filtering -1 IDs, pagination).
  11. How to Select Index Parameters for IVF IndexThis post details best practices for selecting index parameters for the IVF index type in Milvus. It explains the impact of `index_file_size` on raw data and index file sizes, and its relation to memory capacity, recommending values like 1024MB or 2048MB based on performance tests. It also elaborates on `nlist` (number of clusters) and `nprobe` (number of clusters to search) for clustering algorithms, advising on the trade-off between search performance and precision, and suggesting `nlist` as `4 * sqrt(n)` and `nprobe` determined through trial and error.
  12. Accelerating Similarity Search on Really Big Data with Vector IndexingThis post details different IVF index types (FLAT, IVF_FLAT, IVF_SQ8, IVF_SQ8H) and their performance characteristics, providing guidance on index selection based on dataset size, recall requirements, and resource constraints. It includes performance test results for query time and recall rates across various configurations and hardware (CPU/GPU).