chore: change cosine_distance to 1 - e (#10844)

* chore: change cosine_distance to 1 - e

* fix ut

* fix sql logic test

Author: BohuTANG

src/common/vector/src/distance.rs

@@ -30,5 +30,5 @@ pub fn cosine_distance(from: &[f32], to: &[f32]) -> Result<f32> {
     let aa_sum = (&a * &a).sum();
     let bb_sum = (&b * &b).sum();
 
-    Ok((&a * &b).sum() / ((aa_sum).sqrt() * (bb_sum).sqrt()))
+    Ok(1.0 - (&a * &b).sum() / ((aa_sum).sqrt() * (bb_sum).sqrt()))
 }

src/common/vector/tests/it/distance.rs

@@ -21,15 +21,15 @@ fn test_cosine() {
         let y: Vec<f32> = (100..108).map(|v| v as f32).collect();
         let d = cosine_distance(&x, &y).unwrap();
         // from scipy.spatial.distance.cosine
-        approx::assert_relative_eq!(d, 0.900_957);
+        approx::assert_relative_eq!(d, 1.0 - 0.900_957);
     }
 
     {
         let x = vec![3.0, 45.0, 7.0, 2.0, 5.0, 20.0, 13.0, 12.0];
         let y = vec![2.0, 54.0, 13.0, 15.0, 22.0, 34.0, 50.0, 1.0];
         let d = cosine_distance(&x, &y).unwrap();
         // from sklearn.metrics.pairwise import cosine_similarity
-        approx::assert_relative_eq!(d, 0.873_580_6);
+        approx::assert_relative_eq!(d, 1.0 - 0.873_580_6);
     }
 
     {

src/query/functions/tests/it/scalars/testdata/vector.txt

@@ -8,16 +8,16 @@ evaluation:
 | Type   | Float32 | Float32 | Float32 |
 | Domain | {0..=2} | {3..=5} | Unknown |
 | Row 0  | 0       | 3       | NaN     |
-| Row 1  | 1       | 4       | 1       |
-| Row 2  | 2       | 5       | 1       |
+| Row 1  | 1       | 4       | 0       |
+| Row 2  | 2       | 5       | 0       |
 +--------+---------+---------+---------+
 evaluation (internal):
 +--------+----------------------+
 | Column | Data                 |
 +--------+----------------------+
 | a      | Float32([0, 1, 2])   |
 | b      | Float32([3, 4, 5])   |
-| Output | Float32([NaN, 1, 1]) |
+| Output | Float32([NaN, 0, 0]) |
 +--------+----------------------+
 
 

tests/sqllogictests/suites/query/02_function/02_0063_function_vector

@@ -2,4 +2,4 @@
 query F
 select cosine_distance([3.0, 45.0, 7.0, 2.0, 5.0, 20.0, 13.0, 12.0], [2.0, 54.0, 13.0, 15.0, 22.0, 34.0, 50.0, 1.0]) as sim
 ----
-0.8735807
+0.1264193

website/blog/2023-03-24-databend-weekly-86.md

@@ -87,7 +87,7 @@ Databend has added a new function called `cosine_distance`. This function accept
 ```sql
 select cosine_distance([3.0, 45.0, 7.0, 2.0, 5.0, 20.0, 13.0, 12.0], [2.0, 54.0, 13.0, 15.0, 22.0, 34.0, 50.0, 1.0]) as sim
 ----
-0.8735807
+0.1264193
 ```
 
 The Rust implementation efficiently performs calculations by utilizing the `ArrayView` type from the [ndarray](https://crates.io/crates/ndarray) crate.