Vector Database for Similarity Search

May 2025Why AI Databases Don't Need SQLThis post argues for the decline of SQL in AI applications, advocating for natural language and AI agent interfaces for databases. It contrasts traditional relational databases with purpose-built vector databases like Milvus, highlighting Milvus's advantages in native multimodal support, agent integration, semantic intelligence, and optimized relevance. It critiques the performance and philosophical incompatibility of bolting vector capabilities onto SQL databases, citing benchmark data showing Milvus's superior latency and throughput compared to PostgreSQL with pgvector.

Mar 2025What Exactly is a Vector Database and How Does It WorkThis post provides a foundational explanation of what a vector database is and how it works. It details the core concepts of vector embeddings, their role in bridging the semantic gap, and the mathematical principles behind vector similarity, including vector spaces and distance metrics (Euclidean, Cosine, Dot Product, Manhattan). It outlines the typical four-tier architecture of a production vector database (Storage, Index, Query, Service layers) and describes the end-to-end vector search workflow, from data transformation to retrieval. The post also contrasts vector databases with traditional relational databases and vector index libraries, highlighting the limitations of the latter and the comprehensive data management capabilities offered by vector databases.

Jun 2022Scalable and Blazing Fast Similarity Search with Milvus Vector DatabaseThis post introduces the fundamental concepts of unstructured data, its challenges (storage, representation, querying), and the role of embeddings and vector similarity search. It explains what a vector database is, its key requirements (scalability, reliability, speed), and introduces Milvus as an open-source vector database management platform. The post outlines Milvus's capabilities in storing, indexing, and querying billions of embedding vectors, supporting CRUD operations, and attribute filtering. It also provides a blueprint for a visual image search use case.

Mar 2022Building a Vector Database for Scalable Similarity SearchThis post details the thinking process and design principles behind building Milvus, an open-source, cloud-native vector database for scalable similarity search. It explains the need for handling unstructured data by converting it into vectors and the role of similarity search. It differentiates Milvus from a vector search engine by highlighting its database features like sharding, partitioning, replication, and load balancing. The post emphasizes a cloud-native first approach, inspired by shared storage architectures, and a microservices design managed by Kubernetes. Key design principles for Milvus 2.0 include 'Log as data' (using a log broker for all operations), 'Duality of table and log' (aggregating logs into segments), and 'Log persistency' through a pub/sub system. The architecture is presented in four layers: access, coordination, data, and index, with details on the proxy's role in the access layer and the separation of data and control flow. It also discusses standalone and cluster deployment modes.

Dec 2021Milvus in IP Protection：Building a Trademark Similarity Search System with MilvusThis post details the process of building an image-based trademark similarity search system using Milvus. It outlines the steps involved: preparing a dataset of logos, training an image feature extraction model (specifically VGG16), converting logos into vectors using the trained model, and storing these vectors in Milvus for similarity search. The post explains the VGG16 architecture and how Milvus can be used to store and search these vectors, including the concept of creating collections for different product categories.

Nov 2021Frustrated with New Data? Our Vector Database can HelpThis post introduces the fundamental concepts and challenges of building a general-purpose vector database. It discusses the limitations of traditional data analytics for unstructured data and the emergence of the vector data layer through embedding technology. Key attributes of a vector database are outlined: high-efficiency vector operators, support for vector indexing, consistent user experience across deployment environments, support for hybrid search, and a cloud-native architecture. The post also provides a high-level overview of the Milvus 2.0 architecture, emphasizing 'log as data', unified batch and stream processing, stateless components, and microservices.

Oct 2021Building a Video Analysis System with Milvus Vector DatabaseThis post details the construction of a video analysis system using Milvus. It outlines the process of extracting frames from videos using OpenCV, performing object detection with YOLOv3, extracting feature vectors from detected objects using ResNet-50, and finally utilizing Milvus for vector similarity search. The post provides code snippets for frame extraction, object detection, feature extraction, and Milvus search operations, demonstrating a practical application of Milvus for analyzing unstructured video content.

Sep 2021DNA Sequence Classification based on MilvusThis post demonstrates the application of Milvus for DNA sequence classification. It details the process of converting DNA sequences into k-mers, vectorizing these k-mers using CountVectorizer, and then storing and searching these vectors in Milvus. The post highlights Milvus's ability to manage unstructured data and perform high-efficiency retrieval using ANN algorithms, enabling millisecond-level gene classification with higher accuracy than traditional ML classifiers. It also shows how Milvus can be used to calculate genetic similarity between organisms by comparing DNA sequence vectors.

Jul 2021Building a Personalized Product Recommender System with Vipshop and MilvusThis post details Vipshop's implementation of a personalized product recommender system using Milvus. It describes the system architecture, including a write process that synchronizes item feature vectors from MySQL to Milvus via ETL, and a read process that retrieves user preference vectors and queries Milvus for similar item vectors. The post highlights the challenges faced with Elasticsearch for similarity calculations (long response times, high index maintenance costs) and how Milvus addressed these with its distributed deployment and efficient querying. It also details data update strategies, including batch writing to multiple partitions, index building and pre-loading, and using Redis for seamless alias switching between collections. The data recall process involves concurrent querying of Milvus partitions and merging results, achieving an average latency of 30ms. Lessons learned include the benefits of read-write separation, building a connection pool for the Java client, index warm-up for slow queries, tuning nlist and nprobe parameters, and the efficiency of batch import for static data.

Jun 2021Building an Intelligent News Recommendation System Inside Sohu News AppThis post details how Sohu News implemented a news recommendation system using Milvus. Key technical contributions include: 1. Encoding news articles into 968-dimensional semantic vectors by splicing a summary embedding with topic vectors. 2. Encoding user profiles into 968-dimensional semantic vectors by generating tag phrases, identifying relevant topics, and splicing them into tag semantic vectors. 3. Optimizing semantic vector extraction time from days to hours by storing existing tag phrase semantic vectors and using MinHash with Jaccard distance to retrieve similar vectors for new tag phrases. 4. Resolving short text classification issues by using Milvus to query for semantically similar long news articles and applying a voting mechanism to correct misclassifications, achieving over 95% accuracy.

May 2021Accelerating AI in Finance with Milvus, an Open-Source Vector DatabaseThis post introduces Milvus as an open-source vector database specifically for managing vector data, enabling AI application development and MLOps. It details common applications in finance, including enhancing customer experience with banking chatbots, boosting sales with recommender systems, and analyzing unstructured financial data with semantic text mining. It links to demos and further resources for each application.

Apr 2021Making with Milvus Detecting Android Viruses in Real Time for Trend MicroThis post details how Trend Micro uses Milvus for real-time Android virus detection. It describes their system architecture, which involves converting APK Thash values into binary vectors and storing them in Milvus for similarity search. The post compares Milvus to other solutions like Faiss and Elasticsearch plugins, highlighting Milvus's performance advantages. It also details Trend Micro's specific implementation using Hamming distance, a Redis caching layer, and Mishards for sharding on Kubernetes, along with their monitoring setup using Prometheus and Grafana. The system achieves end-to-end query latency under 95ms and fast insertion rates.

Apr 2021Build Semantic Search at SpeedThis post details the integration of Milvus as a vector search engine within Lucidworks' Fusion platform to achieve high-performance semantic search. It outlines a three-stage workflow: ML model encoding of text into vectors, Milvus for efficient vector similarity search, and Solr for retrieving full documents based on Milvus's results. The post highlights Milvus's use of FAISS and its ability to handle large-scale vector datasets. Scale testing results are presented, demonstrating 1500 RPS with an average response time of 40 milliseconds using Milvus with FLAT and HNSW indexes.

Apr 2021How to Make 4 Popular AI Applications with MilvusThis post demonstrates the practical application of Milvus as a vector database by showcasing four distinct AI application examples: NLP chatbots, reverse image search, audio search, and video object detection. It details the integration of Milvus with specific ML models (BERT, VGG, PANNs, YOLOv3, ResNet50) for feature vector extraction and similarity search. For each application, it outlines the 'How to use' steps and the underlying 'How it works' architecture, including data processing and search mechanisms. It also provides links to source code repositories and demo videos.

Mar 2021Making With Milvus AI-Infused Proptech for Personalized Real Estate SearchThis post details how Beike uses Milvus for personalized real estate search. It explains the concept of vector similarity search and how unstructured data is converted into feature vectors. Milvus is used to store and index these vectors, enabling similarity searches based on user queries. The example highlights a specific application in proptech, where property listings are vectorized and searched for similar properties, achieving an average query time of 113ms on a 3 million vector dataset.

Mar 2021Extracting Event Highlights Using iYUNDONG Sports AppThis post details how iYUNDONG integrated Milvus into their sports app to build an intelligent image retrieval system. The system extracts facial features from uploaded photos, converts them into 128-dimensional vectors, and stores them in Milvus for similarity search. Milvus is used for its ability to perform fast, large-scale vector searches and reduce RAM usage, which helped iYUNDONG overcome high operating costs associated with storing billions of feature vectors. The post also touches upon iYUNDONG's infrastructure choices, including cloud storage (AWS S3, Alibaba Cloud OSS), sharding middleware, and Redis caching for scalability and performance.

Feb 2021How we used semantic search to make our search 10x smarterThis post details Tokopedia's adoption of Milvus for semantic search to enhance product relevance. It contrasts keyword-based search (Elasticsearch) with vector representation and similarity search. Milvus was chosen over FAISS due to its ease of use, comprehensive index support, and detailed documentation. A clustered Milvus service with Mishards was deployed for high availability, utilizing Ansible for infrastructure orchestration. The post also explains the role of vector indexing in accelerating similarity search.

Jan 2021Milvus Was Built for Massive-Scale (Think Trillion) Vector Similarity SearchThis post introduces Milvus as a vector data management solution built for massive-scale vector similarity search. It explains the core architectural choices that enable this scale: the use of LSM trees for efficient dynamic data management (handling high insert/delete rates), optimized data insertion and flushing to disk via WAL and mutable buffers, and segment merging to limit fragmentation. It also highlights the importance of indexing for accelerating similarity search, mentioning brute-force search as a baseline and the use of ANN algorithms. Metadata management using OLTP databases (SQLite/MySQL) is also discussed.

Jan 2021Milvus Is an Open-Source Scalable Vector DatabaseThis post introduces Milvus as an open-source scalable vector database. It explains the challenge of searching large unstructured datasets and how AI converts unstructured data into vectors for similarity search. It highlights the need for Approximate Nearest Neighbor (ANN) algorithms to overcome the time-consuming nature of full dataset searches. Milvus is presented as a unified platform that integrates tools like Faiss, NMSLIB, and Annoy to simplify building ML applications. The post emphasizes Milvus's flexibility, distributed architecture for scalability, and compatibility with various hardware and software.

Jan 2021Vector Similarity Search Hides in Plain ViewThis post introduces the concept of vector similarity search and its broad applications across various domains like e-commerce, security, recommendation engines, chatbots, and chemical similarity search. It explains how unstructured data is converted into feature vectors (embeddings) and how approximate nearest neighbor (ANN) algorithms are used to accelerate similarity search. The post also highlights Milvus as an open-source vector data management platform and lists public datasets for AI/ML projects.

Dec 2020Building a Graph-based Recommendation System with Milvus, PinSage, DGL, and MovieLens DatasetsThis post details the process of building a graph-based recommendation system using Milvus, PinSage, and DGL. It outlines data preprocessing steps using MovieLens datasets, training a PinSage model to generate movie embeddings, loading these embeddings into Milvus, and performing vector similarity searches to identify similar movies. The workflow demonstrates how Milvus can be integrated into a recommendation system pipeline for retrieving relevant information based on vector similarity.

Nov 2020Thanks to Milvus, Anyone Can Build a Vector Database for 1+ Billion ImagesThis post introduces Milvus as a solution for building a vector database capable of handling over 1 billion images. It explains how AI (neural networks) converts unstructured data into feature vectors (embeddings) and discusses the limitations of existing vector data management approaches, such as in-memory storage, static data handling, lack of advanced query processing, and insufficient heterogeneous computing optimizations. Milvus is presented as a platform that overcomes these limitations by offering flexible application interfaces, multiple vector index types, dynamic data handling via an LSM tree, and optimizations for CPU and GPU architectures. It claims to achieve high recall rates and fast query times for large datasets, enabling applications like reverse image search, video recommendation systems, and NLP.

Aug 20204 Steps to Building a Video Search SystemThis post details the architecture and steps for building a video search system using Milvus. It outlines the system overview, data preparation, deployment steps (including Docker image building and environment configuration with docker-compose), starting the system, and importing videos. The system uses OpenCV for frame extraction, VGG for feature vector generation, Milvus for vector storage and similarity search, Minio for original video storage, and Redis for mapping correlations between videos and vectors. The post provides specific commands and configuration examples for setting up the system.

Aug 2020The Journey to Optimizing Billion-scale Image Search (1/2)This post details the first generation of an image search system for Yupoo Picture Manager, which manages tens of billions of pictures. The system uses the Perceptual hash (pHash) algorithm for feature extraction and Hamming distance for similarity calculation. To handle large-scale similarity calculations, the pHash values were segmented, and Elasticsearch was used for equivalence matching. The limitations of this approach, such as sensitivity to image integrity changes, are discussed, setting the stage for a second-generation system with different underlying technology.

What this trail covers