From ef07fcb30843724bd98fc0e1b4fdd4f061c32924 Mon Sep 17 00:00:00 2001
From: Roman Rizzi <roman@discourse.org>
Date: Tue, 26 Nov 2024 15:54:20 -0300
Subject: [PATCH] FIX: Skip records without content to classify (#960)

---
 app/jobs/scheduled/embeddings_backfill.rb     |  1 +
 lib/embeddings/vector_representations/base.rb | 25 +++++++++++--------
 .../multilingual_e5_large.rb                  |  8 +++---
 .../vector_rep_shared_examples.rb             |  4 +++
 4 files changed, 24 insertions(+), 14 deletions(-)

diff --git a/app/jobs/scheduled/embeddings_backfill.rb b/app/jobs/scheduled/embeddings_backfill.rb
index db0828f2..c7af4e9f 100644
--- a/app/jobs/scheduled/embeddings_backfill.rb
+++ b/app/jobs/scheduled/embeddings_backfill.rb
@@ -70,6 +70,7 @@ module Jobs
         Post
           .joins("LEFT JOIN #{table_name} ON #{table_name}.post_id = posts.id")
           .where(deleted_at: nil)
+          .where(post_type: Post.types[:regular])
           .limit(limit - rebaked)
 
       # First, we'll try to backfill embeddings for posts that have none
diff --git a/lib/embeddings/vector_representations/base.rb b/lib/embeddings/vector_representations/base.rb
index e1f3ff49..7055ea23 100644
--- a/lib/embeddings/vector_representations/base.rb
+++ b/lib/embeddings/vector_representations/base.rb
@@ -61,18 +61,21 @@ module DiscourseAi
 
           embedding_gen = inference_client
           promised_embeddings =
-            relation.map do |record|
-              materials = { target: record, text: prepare_text(record) }
+            relation
+              .map do |record|
+                prepared_text = prepare_text(record)
+                next if prepared_text.blank?
 
-              Concurrent::Promises
-                .fulfilled_future(materials, pool)
-                .then_on(pool) do |w_prepared_text|
-                  w_prepared_text.merge(
-                    embedding: embedding_gen.perform!(w_prepared_text[:text]),
-                    digest: OpenSSL::Digest::SHA1.hexdigest(w_prepared_text[:text]),
-                  )
-                end
-            end
+                Concurrent::Promises
+                  .fulfilled_future({ target: record, text: prepared_text }, pool)
+                  .then_on(pool) do |w_prepared_text|
+                    w_prepared_text.merge(
+                      embedding: embedding_gen.perform!(w_prepared_text[:text]),
+                      digest: OpenSSL::Digest::SHA1.hexdigest(w_prepared_text[:text]),
+                    )
+                  end
+              end
+              .compact
 
           Concurrent::Promises
             .zip(*promised_embeddings)
diff --git a/lib/embeddings/vector_representations/multilingual_e5_large.rb b/lib/embeddings/vector_representations/multilingual_e5_large.rb
index 605ec8b5..fe611ec9 100644
--- a/lib/embeddings/vector_representations/multilingual_e5_large.rb
+++ b/lib/embeddings/vector_representations/multilingual_e5_large.rb
@@ -81,11 +81,13 @@ module DiscourseAi
         end
 
         def prepare_text(record)
-          if inference_client.class.name.include?("DiscourseClassifier")
-            return "query: #{super(record)}"
+          prepared_text = super(record)
+
+          if prepared_text.present? && inference_client.class.name.include?("DiscourseClassifier")
+            return "query: #{prepared_text}"
           end
 
-          super(record)
+          prepared_text
         end
       end
     end
diff --git a/spec/lib/modules/embeddings/vector_representations/vector_rep_shared_examples.rb b/spec/lib/modules/embeddings/vector_representations/vector_rep_shared_examples.rb
index fce9f612..075e6930 100644
--- a/spec/lib/modules/embeddings/vector_representations/vector_rep_shared_examples.rb
+++ b/spec/lib/modules/embeddings/vector_representations/vector_rep_shared_examples.rb
@@ -79,6 +79,10 @@ RSpec.shared_examples "generates and store embedding using with vector represent
       expect(vector_rep.topic_id_from_representation(expected_embedding_1)).to eq(topic.id)
       expect(vector_rep.topic_id_from_representation(expected_embedding_1)).to eq(topic.id)
     end
+
+    it "does nothing if passed record has no content" do
+      expect { vector_rep.gen_bulk_reprensentations([Topic.new]) }.not_to raise_error
+    end
   end
 
   describe "#asymmetric_topics_similarity_search" do