feat: Include codes with feedback, even if not most similar

This allows for weighting codes higher that may not have been included
in the original similarity query.

lib/medical_transcription/coding.ex

@@ -53,8 +53,11 @@ defmodule MedicalTranscription.Coding do
       MedicalTranscription.Feedback.find_related_feedback(search_vector_for_db, opts)
 
     code_vectors = find_similar(search_vector_for_db, k)
+    code_vectors_for_feedback = find_for_feedback(search_vector_for_db, past_feedbacks)
 
     code_vectors
+    |> Enum.concat(code_vectors_for_feedback)
+    |> Enum.uniq_by(& &1.id)
     |> weight_code_vectors(past_feedbacks)
     |> filter_below_threshold(similarity_threshold)
     |> sort_by_similarity()
@@ -131,4 +134,22 @@ defmodule MedicalTranscription.Coding do
         }
     )
   end
+
+  # Retrieves code vectors referenced in past feedback for a given search text. This allows including additional codes
+  # for the weighting process outside of the closest 5 returned by `find_similar/2`.
+  defp find_for_feedback(search_vector, past_feedbacks) do
+    code_vector_ids_for_past_feedback = Enum.map(past_feedbacks, & &1.code_vector_id)
+
+    Repo.all(
+      from v in CodeVector,
+        order_by: cosine_distance(v.description_vector, ^search_vector),
+        where: v.id in ^code_vector_ids_for_past_feedback,
+        select: %CodeVectorMatch{
+          id: v.id,
+          code: v.code,
+          description: v.description,
+          cosine_similarity: 1 - cosine_distance(v.description_vector, ^search_vector)
+        }
+    )
+  end
 end