[VI] MySQL Native Vector Index : Introduction

Author: mayprasaCreated Aug 4, 2026Updated Sep 10, 2026

Goal

The goal of this project is to implement native vector indexing in MySQL, enabling efficient similarity search while preserving MySQL’s transactional, reliability, operational, and extensibility guarantees.

Executive summary

The Vector Index project will evolve MySQL’s existing VECTOR data type into a complete, native vector storage and retrieval capability. While MySQL can currently store vector values, these columns cannot be indexed and the Community Edition does not expose the similarity functions required for practical AI retrieval workloads.

The project will introduce native VECTOR indexes, vector distance and similarity functions such as cosine similarity, dot product, and Euclidean distance, and approximate nearest-neighbor search using algorithms such as IVF or HNSW. Vector operations will integrate with InnoDB’s transaction, MVCC, redo, undo, crash recovery, replication, backup, security, and operational tooling.

This capability will allow applications to store business data and embeddings in the same transactional system. Source rows and their vectors will commit or roll back together, eliminating synchronization problems commonly introduced when MySQL is paired with an external vector database.

The expected business and technical benefits include:

A unified platform for transactional application data and AI retrieval.
Strong consistency between source records and vector indexes.
ACID-compliant vector updates and crash recovery.
Low-latency similarity search over high-dimensional data.
Reuse of existing MySQL access control, backup, replication, monitoring, and compliance processes.
Reduced application complexity, infrastructure footprint, and operational cost.

The initial delivery should establish distance functions, native ANN query execution, and a production-ready vector index integrated with InnoDB. The architecture should remain extensible so additional indexing strategies and compression techniques can be introduced without changing SQL-level behavior or core transactional guarantees.