Remove redundant allocations in distance function

crates/index/src/hnsw/search.rs

@@ -23,7 +23,7 @@ impl HnswIndex {
         let mut current = ep;
         loop {
             let cur_vec = self.get_vec(current)?;
-            let mut best_score = distance(&query.to_vec(), cur_vec, self.similarity);
+            let mut best_score = distance(query, cur_vec, self.similarity);
             let mut best_id = current;
 
             let empty: &[PointId] = &[];
@@ -46,7 +46,7 @@ impl HnswIndex {
                     continue;
                 }
                 let n_vec = self.get_vec(n)?;
-                let score = distance(&query.to_vec(), n_vec, self.similarity);
+                let score = distance(query, n_vec, self.similarity);
                 if score < best_score {
                     best_score = score;
                     best_id = n;
@@ -91,7 +91,7 @@ impl HnswIndex {
                 .unwrap_or(ep),
         };
 
-        let ep_score = distance(&query.to_vec(), self.get_vec(seed)?, self.similarity);
+        let ep_score = distance(query, self.get_vec(seed)?, self.similarity);
         candidates.push((Reverse(OrdF32::new(ep_score)), seed));
         w_heap.push((OrdF32::new(ep_score), seed));
         visited.insert(seed);
@@ -125,7 +125,7 @@ impl HnswIndex {
                 }
 
                 visited.insert(n);
-                let score = distance(&query.to_vec(), self.get_vec(n)?, self.similarity);
+                let score = distance(query, self.get_vec(n)?, self.similarity);
                 let score = OrdF32::new(score);
                 candidates.push((Reverse(score), n));
                 if w_heap.len() < ef_construction {

crates/index/src/lib.rs

@@ -21,41 +21,38 @@ pub trait VectorIndex: Send + Sync {
 }
 
 /// Distance function to get the distance between two vectors (taken from old version)
-pub fn distance(a: &DenseVector, b: &DenseVector, dist_type: Similarity) -> f32 {
+pub fn distance(a: &[f32], b: &[f32], dist_type: Similarity) -> f32 {
     assert_eq!(a.len(), b.len());
     match dist_type {
-        Similarity::Euclidean => {
-            let score: Vec<f32> = a
-                .iter()
-                .zip(b.iter())
-                .map(|(&x, &y)| (x - y) * (x - y))
-                .collect();
-            score.iter().sum::<f32>().sqrt()
-        }
-        Similarity::Manhattan => {
-            let score: Vec<f32> = a
-                .iter()
-                .zip(b.iter())
-                .map(|(&x, &y)| (x - y).abs())
-                .collect();
-            score.iter().sum::<f32>()
-        }
-        Similarity::Hamming => {
-            let score: Vec<f32> = a
-                .iter()
-                .zip(b.iter())
-                .map(|(&x, &y)| if (x - y).abs() > 1e-8 { 1f32 } else { 0f32 })
-                .collect();
-            score.iter().sum::<f32>()
-        }
+        Similarity::Euclidean => a
+            .iter()
+            .zip(b.iter())
+            .map(|(&x, &y)| {
+                let d = x - y;
+                d * d
+            })
+            .sum::<f32>()
+            .sqrt(),
+        Similarity::Manhattan => a
+            .iter()
+            .zip(b.iter())
+            .map(|(&x, &y)| (x - y).abs())
+            .sum::<f32>(),
+        Similarity::Hamming => a
+            .iter()
+            .zip(b.iter())
+            .map(|(&x, &y)| if (x - y).abs() > 1e-8 { 1f32 } else { 0f32 })
+            .sum::<f32>(),
         Similarity::Cosine => {
-            let p_score: Vec<f32> = a.iter().zip(b.iter()).map(|(&x, &y)| x * y).collect();
-            let p = p_score.iter().sum::<f32>();
-            let q_score: Vec<f32> = a.iter().map(|&n| n * n).collect();
-            let q = q_score.iter().sum::<f32>().sqrt();
-            let r_score: Vec<f32> = b.iter().map(|&n| n * n).collect();
-            let r = r_score.iter().sum::<f32>().sqrt();
-            1.0 - p / (q * r)
+            let mut dot = 0.0f32;
+            let mut norm_a = 0.0f32;
+            let mut norm_b = 0.0f32;
+            for (&x, &y) in a.iter().zip(b.iter()) {
+                dot += x * y;
+                norm_a += x * x;
+                norm_b += y * y;
+            }
+            1.0 - dot / (norm_a.sqrt() * norm_b.sqrt())
         }
     }
 }