PouryaKit.AI.Postgres
Text stored by its embedding and searched by meaning with pgvector.
Text stored by its embedding and searched by meaning with pgvector. A command stores text under a key through IVectorStore, in the Module's own Unit of Work, so a rolled-back command stores nothing, and a query finds the nearest passages, always within the Operation's tenant.
Install
dotnet add package PouryaKit.AI.Postgres --prereleaseWhen to use it
- You find stored text again by what it means.
- A model looks passages up through a Gateway tool.
- You embed text inside a command and want it to commit with the command.
It builds on PouryaKit.AI, which registers the embedding generator, and PouryaKit.Persistence.Postgres, whose Unit of Work it writes through.
Quick start
Add the vector store to a Module stored with UsePouryaKitPostgres, and create its collection in a migration, as the Kit's tests do. The Host registers its embedding generator with AddPouryaKitEmbeddings:
module.AddPouryaKitVectorStore();
migrationBuilder.CreateVectorCollection(VectorsModule.Passages, VectorsModule.Dimensions);A command stores text under a key, then searches by meaning what it stored:
internal sealed class StoreThenSearchPassageCommandHandler(IVectorStore vectors)
: CommandHandler<StoreThenSearchPassageCommand, IReadOnlyList<VectorMatch>>
{
public override async Task<Result<IReadOnlyList<VectorMatch>>> Handle(
StoreThenSearchPassageCommand command,
CancellationToken cancellationToken
)
{
await vectors.StoreAsync(
VectorsModule.Passages,
command.Key,
command.Text,
cancellationToken
);
return Result.Success(
await vectors.SearchAsync(VectorsModule.Passages, command.Text, 1, cancellationToken)
);
}
}What you get
- Each row keyed by the Operation's tenant: a call never passes a tenant.
- A store that commits or rolls back with the command.
- In a command, a search that sees what the command stored.
Storing calls the model inside the Unit of Work's transaction, so a slow model holds the Operation's locks.