diff --git a/web/packages/new/photos/services/ml/clip.ts b/web/packages/new/photos/services/ml/clip.ts
index 4981a25906..e1f14cdde6 100644
--- a/web/packages/new/photos/services/ml/clip.ts
+++ b/web/packages/new/photos/services/ml/clip.ts
@@ -178,7 +178,8 @@ export const clipMatches = async (
     const textEmbedding = normalized(t);
     const items = (await cachedOrReadCLIPIndexes()).map(
         ({ fileID, embedding }) =>
-
+            // The dot product gives us cosine similarity here since both the
+            // vectors are already normalized.
             [fileID, dotProduct(embedding, textEmbedding)] as const,
     );
     // This score threshold was obtain heuristically. 0.2 generally gives solid
diff --git a/web/packages/new/photos/services/ml/math.ts b/web/packages/new/photos/services/ml/math.ts
index ca965da0a5..9e78dcd822 100644
--- a/web/packages/new/photos/services/ml/math.ts
+++ b/web/packages/new/photos/services/ml/math.ts
@@ -24,7 +24,7 @@ export const clamp = (value: number, min: number, max: number) =>
  * product. When we're computing dot products in a hot loop, skipping over that
  * unnecessary renormalization matters.
  *
- * When comparing embeddings we usually want is the cosine similarity, but when
+ * When comparing embeddings we usually want is the cosine similarity, but if
  * both the embeddings involved are already normalized, we can save the norm
  * calculations and directly do their `dotProduct`.
  *