等待 SQL:202y: 向量

发布于 8 月 3 日

Summary: Since SQL:2023 was published, work on the SQL standard continues. Vectors have become popular in databases related to LLM and "AI" use cases. SQL now has a new vector data type with two arguments (dimension count and coordinate type). Utility functions like vector_dimension_count(), vector_norm(), and vector_serialize() are available. Vectors can be compared for similarity using functions like cosine, dot, etc. in vector_distance(). Usually, results are ordered by vector distance and a top-N limit is applied. Approximate results are now supported with the FETCH APPROX FIRST clause, and an additional extension allows specifying a range. Indexing for vectors is not part of the standard and is up to the implementation.
Main Points:
- SQL standard progress after SQL:2023.
- Introduction of vector data type.
- Operations on vector type with various functions.
- Comparison and sorting by vector distance.
- Use of FETCH APPROX FIRST for approximate results.
- Extension for specifying a range of approximate results.
Key Information:
- Naming convention for next SQL standard as SQL:202y.
- vector type arguments and example usage.
- Different vector similarity calculation functions.
- Default use of APPROX with vector type.
Important Details:
- How to insert data into vector type.
- Explanation of how to use vector_distance() with different functions.
- Illustration of applying top-N limit and approximate range.
- Note that indexing for vectors is implementation-dependent.

阅读 12