The Semantic Kernel Vector Store data architecture (Experimental)

Vector Store abstractions in Semantic Kernel are based on three main components: vector stores, collections and records. Records are contained by collections, and collections are contained by vector stores.

• A vector store maps to an instance of a database
• A collection is a collection of records including any index required to query or filter those records
• A record is an individual data entry in the database

Collections in different databases

The underlying implementation of what a collection is, will vary by connector and is influenced by how each database groups and indexes records. Most databases have a concept of a collection of records and there is a natural mapping between this concept and the Vector Store abstraction collection. Note that this concept may not always be referred to as a collection in the underlying database.