👁 Image
Microsoft.KernelMemory.MemoryDb.Postgres 0.98.250508.3

Kernel Memory with Postgres + pgvector

👁 Nuget package
👁 Discord

This project contains the Postgres adapter allowing to use Kernel Memory with Postgres+pgvector.

SELECT * FROM pg_extension

CREATE EXTENSION vector

To use Postgres with Kernel Memory:

1. Have a PostgreSQL instance ready, e.g. checkout Azure Database for PostgreSQL

2. Verify your Postgres instance supports vectors, e.g. run SELECT * FROM pg_extension

3. Add Postgres connection string to appsettings.json (or appsettings.Development.json), for example:

   {
    "KernelMemory": {
    "Services": {
    "Postgres": {
    "ConnectionString": "Host=localhost;Port=5432;Username=myuser;Password=mypassword;Database=mydatabase"
    }
    }
    }
   }
   

4. Configure KM builder to store memories in Postgres, and to persist documents, for example:

   // using Microsoft.KernelMemory;
   // using Microsoft.KernelMemory.Postgres;
   // using Microsoft.Extensions.Configuration;
   
   var postgresConfig = new PostgresConfig();
   
   new ConfigurationBuilder()
    .AddJsonFile("appsettings.json")
    .AddJsonFile("appsettings.development.json", optional: true)
    .AddJsonFile("appsettings.Development.json", optional: true)
    .Build()
    .BindSection("KernelMemory:Services:Postgres", postgresConfig);
   
   var memory = new KernelMemoryBuilder()
    .WithPostgresMemoryDb(postgresConfig)
    .WithSimpleFileStorage(SimpleFileStorageConfig.Persistent)
    .WithAzureOpenAITextGeneration(azureOpenAIConfig)
    .WithAzureOpenAITextEmbeddingGeneration(azureOpenAIConfig)
    .Build();
   

Neighbor search indexes, quality and performance

The connector does not create IVFFlat or HNSW indexes on Postgres tables, and uses exact nearest neighbor search. HNSW (Hierarchical Navigable Small World) has been introduced in pgvector 0.5.0 and is not available in some Postgres instances.

Depending on your scenario you might want to create these indexes manually, considering precision and performance trade-offs, or you can customize the SQL used to create tables via configuration.

SQL to add IVFFlat: CREATE INDEX ON %%table_name%% USING ivfflat (embedding vector_cosine_ops) WITH (lists = 1000);

SQL to add HNSW: CREATE INDEX ON %%table_name%% USING hnsw (embedding vector_cosine_ops);

See https://github.com/pgvector/pgvector for more information.

Memory Indexes and Postgres tables

The Postgres memory connector will create "memory indexes" automatically, one DB table for each memory index.

Table names have a configurable prefix, used to filter out other tables that might be present. The prefix is mandatory, cannot be empty, we suggest using the default km- prefix. Note that the Postgres connector automatically converts _ underscores to - dashes to have a consistent behavior with other storage types supported by Kernel Memory.

Overall we recommend not mixing external tables in the same DB used for Kernel Memory.

Column names and table schema

The connector uses a default schema with predefined columns and indexes.

You can change the field names, and if you need to add additional columns or indexes, you can also customize the CREATE TABLE SQL statement. You can use this approach, for example, to use IVFFlat or HNSW.

See PostgresConfig class for more details.

Here's an example where PostgresConfig is stored in appsettings.json and the table schema is customized, with custom names and additional fields.

The SQL statement requires two special placeholders:

• %%table_name%%: replaced at runtime with the table name
• %%vector_size%%: replaced at runtime with the embedding vectors size

There's a third optional placeholder we recommend using, to better handle concurrency, e.g. in combination with pg_advisory_xact_lock (exclusive transaction level advisory locks):

• %%lock_id%%: replaced at runtime with a number

Also:

• TableNamePrefix is mandatory string added to all KM tables
• Columns is a required map describing where KM will store its data. If you have additional columns you don't need to list them here, only in SQL statement.
• CreateTableSql is your optional custom SQL statement used to create tables. The column names must match those used in Columns.

{
 "KernelMemory": {
 "Services": {
 "Postgres": {

 "TableNamePrefix": "memory_",

 "Columns": {
 "id": "_pk",
 "embedding": "embedding",
 "tags": "labels",
 "content": "chunk",
 "payload": "extras"
 },

 "CreateTableSql": [
 "BEGIN; ",
 "SELECT pg_advisory_xact_lock(%%lock_id%%); ",
 "CREATE TABLE IF NOT EXISTS %%table_name%% ( ",
 " _pk TEXT NOT NULL PRIMARY KEY, ",
 " embedding vector(%%vector_size%%), ",
 " labels TEXT[] DEFAULT '{}'::TEXT[] NOT NULL, ",
 " chunk TEXT DEFAULT '' NOT NULL, ",
 " extras JSONB DEFAULT '{}'::JSONB NOT NULL, ",
 " my_field1 TEXT DEFAULT '', ",
 " _update TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP ",
 "); ",
 "CREATE INDEX ON %%table_name%% USING GIN(labels); ",
 "CREATE INDEX ON %%table_name%% USING ivfflat (embedding vector_cosine_ops); ",
 "COMMIT; "
 ]

 }
 }
 }
}