use maxKNNScore

knn_test.go

@@ -20,6 +20,7 @@ package bleve
 import (
 	"archive/zip"
 	"encoding/json"
+	"math"
 	"math/rand"
 	"testing"
 
@@ -355,8 +356,8 @@ func runKNNTest(t *testing.T, randomizeDocuments bool) {
 			numIndexPartitions: 1,
 			expectedResults: map[string]testResult{
 				"doc7": {
-					score:          2357.022603955158,
-					scoreBreakdown: []float64{0, 0, 7071.067811865475},
+					score:          math.MaxFloat64,
+					scoreBreakdown: []float64{0, 0, math.MaxFloat64 / 3.0},
 				},
 				"doc29": {
 					score:          0.6774608026082964,
@@ -402,8 +403,8 @@ func runKNNTest(t *testing.T, randomizeDocuments bool) {
 			numIndexPartitions: 4,
 			expectedResults: map[string]testResult{
 				"doc7": {
-					score:          2357.022603955158,
-					scoreBreakdown: []float64{0, 0, 7071.067811865475},
+					score:          math.MaxFloat64,
+					scoreBreakdown: []float64{0, 0, math.MaxFloat64 / 3.0},
 				},
 				"doc29": {
 					score:          0.567426591648309,

search/scorer/scorer_knn.go

@@ -18,6 +18,7 @@
 package scorer
 
 import (
+	"math"
 	"reflect"
 
 	"github.com/blevesearch/bleve/v2/search"
@@ -60,8 +61,9 @@ func NewKNNQueryScorer(queryVector []float32, queryField string, queryBoost floa
 	}
 }
 
-// TODO: Better value needed here?
-const maxEuclideanDistance = 10000.0
+// Score used when the knnMatch.Score = 0 ->
+// the query and indexed vector are exactly the same.
+const maxKNNScore = math.MaxFloat64
 
 func (sqs *KNNQueryScorer) Score(ctx *search.SearchContext,
 	knnMatch *index.VectorDoc) *search.DocumentMatch {
@@ -70,20 +72,18 @@ func (sqs *KNNQueryScorer) Score(ctx *search.SearchContext,
 	if sqs.includeScore || sqs.options.Explain {
 		var scoreExplanation *search.Explanation
 		score := knnMatch.Score
-		if sqs.similarityMetric == index.EuclideanDistance {
-			// in case of euclidean distance being the distance metric,
-			// an exact vector (perfect match), would return distance = 0
-			if score == 0 {
-				score = maxEuclideanDistance
-			} else {
-				// euclidean distances need to be inverted to work with
-				// tf-idf scoring
-				score = 1.0 / score
-			}
+		// in case of euclidean distance being the distance metric,
+		// an exact vector (perfect match), would return distance = 0
+		if score == 0 {
+			score = maxKNNScore
+		} else {
+			// euclidean distances need to be inverted to work with
+			// tf-idf scoring
+			score = 1.0 / score
 		}
 
 		// if the query weight isn't 1, multiply
-		if sqs.queryWeight != 1.0 {
+		if sqs.queryWeight != 1.0 && score != maxKNNScore {
 			score = score * sqs.queryWeight
 		}