discourse-ai/lib/tasks/modules/embeddings/database.rake

59 lines
2.1 KiB
Ruby

# frozen_string_literal: true
desc "Creates tables to store embeddings"
task "ai:embeddings:create_table" => [:environment] do
DiscourseAi::Database::Connection.db.exec(<<~SQL)
CREATE EXTENSION IF NOT EXISTS vector;
SQL
DiscourseAi::Embeddings::Model.enabled_models.each do |model|
DiscourseAi::Database::Connection.db.exec(<<~SQL)
CREATE TABLE IF NOT EXISTS topic_embeddings_#{model.name.underscore} (
topic_id bigint PRIMARY KEY,
embedding vector(#{model.dimensions})
);
SQL
end
end
desc "Backfill embeddings for all topics"
task "ai:embeddings:backfill", [:start_topic] => [:environment] do
public_categories = Category.where(read_restricted: false).pluck(:id)
topic_embeddings = DiscourseAi::Embeddings::Topic.new
Topic
.where("id >= ?", args[:start_topic] || 0)
.where("category_id IN (?)", public_categories)
.where(deleted_at: nil)
.order(id: :asc)
.find_each do |t|
print "."
topic_embeddings.generate_and_store_embeddings_for(t)
end
end
desc "Creates indexes for embeddings"
task "ai:embeddings:index", [:work_mem] => [:environment] do |_, args|
# Using extension maintainer's recommendation for ivfflat indexes
# Results are not as good as without indexes, but it's much faster
# Disk usage is ~1x the size of the table, so this double table total size
count = Topic.count
lists = count < 1_000_000 ? count / 1000 : Math.sqrt(count).to_i
probes = count < 1_000_000 ? lists / 10 : Math.sqrt(lists).to_i
DiscourseAi::Database::Connection.db.exec("SET work_mem TO '#{args[:work_mem] || "1GB"}';")
DiscourseAi::Embeddings::Model.enabled_models.each do |model|
DiscourseAi::Database::Connection.db.exec(<<~SQL)
CREATE INDEX IF NOT EXISTS
topic_embeddings_#{model.name.underscore}_search
ON
topic_embeddings_#{model.name.underscore}
USING
ivfflat (embedding #{model.pg_index})
WITH
(lists = #{lists});
SQL
end
DiscourseAi::Database::Connection.db.exec("RESET work_mem;")
DiscourseAi::Database::Connection.db.exec("SET ivfflat.probes = #{probes};")
end