fix bug with angular distance and floating point error; added cosine distance

1 files changed, 16 insertions, 1 deletions

diff --git a/src/embedders/vecmetric.rs b/src/embedders/vecmetric.rs
index 65d71df..2ebd170 100644
--- a/src/embedders/vecmetric.rs
+++ b/src/embedders/vecmetric.rs
@@ -6,14 +6,18 @@ pub trait VecMetric: MetricElem + From<Vec<f32>> {}
 #[derive(Decode, Encode)]
 pub struct AngularDistance(pub Vec<f32>);
 #[derive(Decode, Encode)]
+pub struct CosineDistance(pub Vec<f32>);
+#[derive(Decode, Encode)]
 pub struct EuclidianDistance(pub Vec<f32>);
 #[derive(Decode, Encode)]
 pub struct ManhattenDistance(pub Vec<f32>);
 
 impl VecMetric for AngularDistance {}
+impl VecMetric for CosineDistance {}
 impl VecMetric for EuclidianDistance {}
 impl VecMetric for ManhattenDistance {}
 #[rustfmt::skip] impl From<Vec<f32>> for AngularDistance { fn from(value: Vec<f32>) -> Self { Self(value) } }
+#[rustfmt::skip] impl From<Vec<f32>> for CosineDistance { fn from(value: Vec<f32>) -> Self { Self(value) } }
 #[rustfmt::skip] impl From<Vec<f32>> for EuclidianDistance { fn from(value: Vec<f32>) -> Self { Self(value) } }
 #[rustfmt::skip] impl From<Vec<f32>> for ManhattenDistance { fn from(value: Vec<f32>) -> Self { Self(value) } }
 
@@ -28,7 +32,18 @@ impl MetricElem for AngularDistance {
         let mag_a = self.0.iter().map(|x| x.powi(2)).sum::<f32>();
         let mag_b = other.0.iter().map(|x| x.powi(2)).sum::<f32>();
         let cossim = x / (mag_a * mag_b).sqrt();
-        cossim.acos() as f64
+        // clamp is require because floating point errors
+        cossim.clamp(-1., 1.).acos() as f64
+    }
+}
+impl MetricElem for CosineDistance {
+    fn dist(&self, other: &Self) -> f64 {
+        self.0
+            .iter()
+            .zip(other.0.iter())
+            .map(|(a, b)| (*a - *b) * (*b - *a))
+            .sum::<f32>()
+            .sqrt() as f64
     }
 }
 impl MetricElem for EuclidianDistance {