From a355cffc239f55831ae016acb5efd4fdb6d4f933 Mon Sep 17 00:00:00 2001
From: Andrew Kane <andrew@ankane.org>
Date: Tue, 3 Sep 2024 20:25:06 -0700
Subject: [PATCH] Updated exact scores to be consistent with approximate scores

---
 lib/searchkick/query.rb |  7 ++++---
 test/knn_test.rb        | 14 +++-----------
 2 files changed, 7 insertions(+), 14 deletions(-)

diff --git a/lib/searchkick/query.rb b/lib/searchkick/query.rb
index 02c7efca..17622384 100644
--- a/lib/searchkick/query.rb
+++ b/lib/searchkick/query.rb
@@ -927,7 +927,8 @@ def set_knn(payload, knn, per_page, offset)
                   query_value: vector,
                   space_type: space_type
                 }
-              }
+              },
+              boost: distance == "cosine" ? 0.5 : 1.0
             }
           }
         else
@@ -947,9 +948,9 @@ def set_knn(payload, knn, per_page, offset)
           source =
             case distance
             when "cosine"
-              "cosineSimilarity(params.query_vector, params.field) + 1.0"
+              "(cosineSimilarity(params.query_vector, params.field) + 1.0) * 0.5"
             when "euclidean"
-              "1 / (1 + l2norm(params.query_vector, params.field))"
+              "double l2 = l2norm(params.query_vector, params.field); 1 / (1 + l2 * l2)"
             else
               raise ArgumentError, "Unknown distance: #{distance}"
             end
diff --git a/test/knn_test.rb b/test/knn_test.rb
index c3ed506e..fc284c45 100644
--- a/test/knn_test.rb
+++ b/test/knn_test.rb
@@ -20,8 +20,7 @@ def test_basic_exact
     assert_order "*", ["A", "B"], knn: {field: :embedding, vector: [1, 2, 3], exact: true}
 
     scores = Product.search(knn: {field: :embedding, vector: [1, 2, 3], exact: true}).hits.map { |v| v["_score"] }
-    # TODO match approximate
-    assert_in_delta 2, scores[0]
+    assert_in_delta 1, scores[0]
     assert_in_delta 0, scores[1]
   end
 
@@ -72,7 +71,6 @@ def test_euclidean
     assert_order "*", ["A", "B"], knn: {field: :factors, vector: [1, 2, 3]}
 
     scores = Product.search(knn: {field: :factors, vector: [1, 2, 3]}).hits.map { |v| v["_score"] }
-    # TODO return distance
     assert_in_delta 1.0 / (1 + 0), scores[0]
     assert_in_delta 1.0 / (1 + 5**2), scores[1]
   end
@@ -82,13 +80,8 @@ def test_euclidean_exact
     assert_order "*", ["A", "B"], knn: {field: :embedding, vector: [1, 2, 3], distance: "euclidean"}
 
     scores = Product.search(knn: {field: :embedding, vector: [1, 2, 3], distance: "euclidean"}).hits.map { |v| v["_score"] }
-    # TODO return distance
     assert_in_delta 1.0 / (1 + 0), scores[0]
-    if Searchkick.opensearch?
-      assert_in_delta 1.0 / (1 + 5**2), scores[1]
-    else
-      assert_in_delta 1.0 / (1 + 5), scores[1]
-    end
+    assert_in_delta 1.0 / (1 + 5**2), scores[1]
   end
 
   def test_unindexed
@@ -98,8 +91,7 @@ def test_unindexed
     assert_order "*", ["A", "B"], knn: {field: :vector, vector: [1, 2, 3], distance: "cosine"}
 
     scores = Product.search(knn: {field: :vector, vector: [1, 2, 3], distance: "cosine"}).hits.map { |v| v["_score"] }
-    # TODO match approximate
-    assert_in_delta 2, scores[0]
+    assert_in_delta 1, scores[0]
     assert_in_delta 0, scores[1]
 
     error = assert_raises(ArgumentError) do