add method for threshold->internal distance

2 files changed, 28 insertions, 1 deletions

diff --git a/searchlib/src/vespa/searchlib/tensor/distance_function.h b/searchlib/src/vespa/searchlib/tensor/distance_function.h
index 30ad1876317..724f83b6129 100644
--- a/searchlib/src/vespa/searchlib/tensor/distance_function.h
+++ b/searchlib/src/vespa/searchlib/tensor/distance_function.h
@@ -11,15 +11,24 @@ namespace search::tensor {
 /**
  * Interface used to calculate the distance between two n-dimensional vectors.
  *
- * The vectors must be of same size and same type (float or double).
+ * The vectors must be of same size and same cell type (float or double).
  * The actual implementation must know which type the vectors are.
  */
 class DistanceFunction {
 public:
     using UP = std::unique_ptr<DistanceFunction>;
     virtual ~DistanceFunction() {}
+
+    // calculate internal distance (comparable)
     virtual double calc(const vespalib::eval::TypedCells& lhs, const vespalib::eval::TypedCells& rhs) const = 0;
+
+    // convert threshold (external distance units) to internal units
+    virtual double convert_threshold(double threshold) const = 0;
+
+    // convert internal distance to rawscore (1.0 / (1.0 + d))
     virtual double to_rawscore(double distance) const = 0;
+
+    // calculate internal distance, early return allowed if > limit
     virtual double calc_with_limit(const vespalib::eval::TypedCells& lhs,
                                    const vespalib::eval::TypedCells& rhs,
                                    double limit) const = 0;
diff --git a/searchlib/src/vespa/searchlib/tensor/distance_functions.h b/searchlib/src/vespa/searchlib/tensor/distance_functions.h
index 8db7b1f48f1..bccf8c0bdb6 100644
--- a/searchlib/src/vespa/searchlib/tensor/distance_functions.h
+++ b/searchlib/src/vespa/searchlib/tensor/distance_functions.h
@@ -26,6 +26,9 @@ public:
         assert(sz == rhs_vector.size());
         return _computer.squaredEuclideanDistance(&lhs_vector[0], &rhs_vector[0], sz);
     }
+    double convert_threshold(double threshold) const override {
+        return threshold*threshold;
+    }
     double to_rawscore(double distance) const override {
         double d = sqrt(distance);
         double score = 1.0 / (1.0 + d);
@@ -75,6 +78,10 @@ public:
         double distance = 1.0 - cosine_similarity; // in range [0,2]
         return distance;
     }
+    double convert_threshold(double threshold) const override {
+        double cosine_similarity = cos(threshold);
+        return 1.0 - cosine_similarity;
+    }
     double to_rawscore(double distance) const override {
         double cosine_similarity = 1.0 - distance;
         // should be in in range [-1,1] but roundoff may cause problems:
@@ -112,6 +119,9 @@ public:
         double score = 1.0 - _computer.dotProduct(&lhs_vector[0], &rhs_vector[0], sz);
         return std::max(0.0, score);
     }
+    double convert_threshold(double threshold) const override {
+        return threshold;
+    }
     double to_rawscore(double distance) const override {
         double score = 1.0 / (1.0 + distance);
         return score;
@@ -163,6 +173,11 @@ public:
         double hav_central_angle = hav_lat + cos(lat_A)*cos(lat_B)*hav_lon;
         return hav_central_angle;
     }
+    double convert_threshold(double threshold) const override {
+        double half_angle = threshold / (2 * 6371.0088);
+        double rt_hav = sin(half_angle);
+        return rt_hav * rt_hav;
+    }
     double to_rawscore(double distance) const override {
         double hav_diff = sqrt(distance);
         // distance in kilometers:
@@ -197,6 +212,9 @@ public:
         }
         return (double)sum;
     }
+    double convert_threshold(double threshold) const override {
+        return threshold;
+    }
     double to_rawscore(double distance) const override {
         double score = 1.0 / (1.0 + distance);
         return score;