Vector databases are systems designed for the storage, indexing, and retrieval of embedding vectors, excelling at answering "which vectors are most similar to the query." They are commonly used for semantic search, recommendations, deduplication, and multimodal search, and are a common foundational component for Retrieval-Augmented Generation (RAG) applications. Their relationship with traditional databases is not mutually exclusive: some relational databases and search engines already support vector types, similarity queries, or vector indexes.

This field did not emerge from a vacuum: embedding-based retrieval has a long research history, similarity search has an even longer history, and early systems often integrated vector indexes into relational databases or search systems. The Zilliz team introduced a system design for vector data management in the SIGMOD paper "Milvus: A Purpose-Built Vector Data Management System" in 2021. A 2023 review, "Survey of Vector Database Management Systems" surveyed this category and noted that there were already more than 20 commercial vector database systems at the time; this does not imply that all these systems emerged after 2021.

However, whether it is Milvus, which was open-sourced in 2019, or the numerous similar systems that emerged later, they all face a common set of technical challenges. Such systems must overcome five hurdles: the ambiguity of semantic similarity, the large size of vectors, the high cost of similarity computation, the lack of natural partitions, and hybrid queries involving "attribute conditions + vector similarity." Corresponding engineering approaches include using quantization compression to reduce storage and computational overhead, using graph indexes (such as HNSW) to accelerate nearest neighbor searches, and using query optimization and execution engines to handle filtering and aggregation; in large-scale scenarios, systems often adopt Approximate Nearest Neighbor (ANN) indexes, making trade-offs between recall and latency; however, vector retrieval can also use exact search, so the notion that "vector indexes are necessarily approximate" is inaccurate. In 2023, the review by Pan, Wang, and Li, "Survey of Vector Database Management Systems" provided a systematic overview of this category. It is precisely these engineering accumulations that have turned "semantics" from a vague concept into indexable and scalable infrastructure.