FEATURE: Add BGE-M3 embeddings support (#569)

BAAI/bge-m3 is an interesting model, that is multilingual and with a
context size of 8192. Even with a 16x larger context, it's only 4x slower
to compute it's embeddings on the worst case scenario.

Also includes a minor refactor of the rake task, including setting model
and concurrency levels when running the backfill task.

config/settings.yml

@@ -264,6 +264,7 @@ discourse_ai:
       - multilingual-e5-large
       - bge-large-en
       - gemini
+      - bge-m3
     validator: "DiscourseAi::Configuration::EmbeddingsModelValidator"
   ai_embeddings_per_post_enabled:
     default: false

db/migrate/20240410170000_add_embeddings_tablesfor_bge_m3.rb

@@ -0,0 +1,38 @@
+# frozen_string_literal: true
+
+class AddEmbeddingsTablesforBgeM3 < ActiveRecord::Migration[7.0]
+  def change
+    create_table :ai_topic_embeddings_8_1, id: false do |t|
+      t.integer :topic_id, null: false
+      t.integer :model_version, null: false
+      t.integer :strategy_version, null: false
+      t.text :digest, null: false
+      t.column :embeddings, "vector(1024)", null: false
+      t.timestamps
+
+      t.index :topic_id, unique: true
+    end
+    create_table :ai_post_embeddings_8_1, id: false do |t|
+      t.integer :post_id, null: false
+      t.integer :model_version, null: false
+      t.integer :strategy_version, null: false
+      t.text :digest, null: false
+      t.column :embeddings, "vector(1024)", null: false
+      t.timestamps
+
+      t.index :post_id, unique: true
+    end
+    create_table :ai_document_fragment_embeddings_8_1, id: false do |t|
+      t.integer :rag_document_fragment_id, null: false
+      t.integer :model_version, null: false
+      t.integer :strategy_version, null: false
+      t.text :digest, null: false
+      t.column :embeddings, "vector(1024)", null: false
+      t.timestamps
+
+      t.index :rag_document_fragment_id,
+              unique: true,
+              name: "rag_document_fragment_id_embeddings_8_1"
+    end
+  end
+end

lib/embeddings/vector_representations/base.rb

@@ -11,11 +11,12 @@ module DiscourseAi
             [
               DiscourseAi::Embeddings::VectorRepresentations::AllMpnetBaseV2,
               DiscourseAi::Embeddings::VectorRepresentations::BgeLargeEn,
+              DiscourseAi::Embeddings::VectorRepresentations::BgeM3,
               DiscourseAi::Embeddings::VectorRepresentations::Gemini,
               DiscourseAi::Embeddings::VectorRepresentations::MultilingualE5Large,
-              DiscourseAi::Embeddings::VectorRepresentations::TextEmbeddingAda002,
-              DiscourseAi::Embeddings::VectorRepresentations::TextEmbedding3Small,
               DiscourseAi::Embeddings::VectorRepresentations::TextEmbedding3Large,
+              DiscourseAi::Embeddings::VectorRepresentations::TextEmbedding3Small,
+              DiscourseAi::Embeddings::VectorRepresentations::TextEmbeddingAda002,
             ].find { _1.name == model_name }
           end
 

lib/embeddings/vector_representations/bge_m3.rb

@@ -0,0 +1,56 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module Embeddings
+    module VectorRepresentations
+      class BgeM3 < Base
+        class << self
+          def name
+            "bge-m3"
+          end
+
+          def correctly_configured?
+            DiscourseAi::Inference::HuggingFaceTextEmbeddings.configured?
+          end
+
+          def dependant_setting_names
+            %w[ai_hugging_face_tei_endpoint_srv ai_hugging_face_tei_endpoint]
+          end
+        end
+
+        def vector_from(text, asymetric: false)
+          truncated_text = tokenizer.truncate(text, max_sequence_length - 2)
+          DiscourseAi::Inference::HuggingFaceTextEmbeddings.perform!(truncated_text).first
+        end
+
+        def dimensions
+          1024
+        end
+
+        def max_sequence_length
+          8192
+        end
+
+        def id
+          8
+        end
+
+        def version
+          1
+        end
+
+        def pg_function
+          "<#>"
+        end
+
+        def pg_index_type
+          "vector_ip_ops"
+        end
+
+        def tokenizer
+          DiscourseAi::Tokenizer::BgeM3Tokenizer
+        end
+      end
+    end
+  end
+end

lib/tasks/modules/embeddings/database.rake

@@ -1,35 +1,49 @@
 # frozen_string_literal: true
 
 desc "Backfill embeddings for all topics and posts"
-task "ai:embeddings:backfill" => [:environment] do
+task "ai:embeddings:backfill", %i[model concurrency] => [:environment] do |_, args|
   public_categories = Category.where(read_restricted: false).pluck(:id)
 
   strategy = DiscourseAi::Embeddings::Strategies::Truncation.new
-  vector_rep = DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)
+  if args[:model].present?
+    vector_rep =
+      DiscourseAi::Embeddings::VectorRepresentations::Base.find_representation(args[:model]).new(
+        strategy,
+      )
+  else
+    vector_rep =
+      DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)
+  end
   table_name = vector_rep.topic_table_name
 
+  topics =
     Topic
       .joins("LEFT JOIN #{table_name} ON #{table_name}.topic_id = topics.id")
       .where("#{table_name}.topic_id IS NULL")
       .where("category_id IN (?)", public_categories)
       .where(deleted_at: nil)
       .order("topics.id DESC")
-    .find_each do |t|
+
-      print "."
+  Parallel.each(topics.all, in_processes: args[:concurrency].to_i, progress: "Topics") do |t|
+    ActiveRecord::Base.connection_pool.with_connection do
       vector_rep.generate_representation_from(t)
     end
+  end
 
   table_name = vector_rep.post_table_name
+  posts =
     Post
       .joins("LEFT JOIN #{table_name} ON #{table_name}.post_id = posts.id")
       .where("#{table_name}.post_id IS NULL")
       .where(deleted_at: nil)
       .order("posts.id DESC")
-    .find_each do |t|
+
-      print "."
+  Parallel.each(posts.all, in_processes: args[:concurrency].to_i, progress: "Posts") do |t|
+    ActiveRecord::Base.connection_pool.with_connection do
       vector_rep.generate_representation_from(t)
     end
   end
+end
 
 desc "Creates indexes for embeddings"
 task "ai:embeddings:index", [:work_mem] => [:environment] do |_, args|

lib/tokenizer/bge_m3_tokenizer.rb

@@ -0,0 +1,11 @@
+# frozen_string_literal: true
+
+module DiscourseAi
+  module Tokenizer
+    class BgeM3Tokenizer < BasicTokenizer
+      def self.tokenizer
+        @@tokenizer ||= Tokenizers.from_file("./plugins/discourse-ai/tokenizers/bge-m3.json")
+      end
+    end
+  end
+end

spec/shared/tokenizer_spec.rb

@@ -176,3 +176,32 @@ describe DiscourseAi::Tokenizer::BgeLargeEnTokenizer do
     end
   end
 end
+
+describe DiscourseAi::Tokenizer::BgeM3Tokenizer do
+  describe "#size" do
+    describe "returns a token count" do
+      it "for a sentence with punctuation and capitalization and numbers" do
+        expect(described_class.size("Hello, World! 123")).to eq(7)
+      end
+    end
+  end
+
+  describe "#truncate" do
+    it "truncates a sentence" do
+      sentence = "foo bar baz qux quux corge grault garply waldo fred plugh xyzzy thud"
+      expect(described_class.truncate(sentence, 3)).to eq("foo")
+    end
+
+    it "truncates a sentence successfully at a multibyte unicode character" do
+      sentence = "foo bar 👨🏿‍👩🏿‍👧🏿‍👧🏿 baz qux quux corge grault garply waldo fred plugh xyzzy thud"
+      expect(described_class.truncate(sentence, 7)).to eq("foo bar 👨🏿")
+    end
+
+    it "truncates unicode characters properly when they use more than one token per char" do
+      sentence = "我喜欢吃比萨"
+      original_size = described_class.size(sentence)
+      expect(described_class.size(described_class.truncate(sentence, original_size - 2))).to be <
+        original_size
+    end
+  end
+end

tokenizers/README.md

@@ -25,3 +25,7 @@ Licensed under MIT License
 ## mixtral
 
 Licensed under Apache 2.0 License
+
+## bge-m3
+
+Licensed under MIT License