Normalize vectors in advance for cosine

CHANGELOG.md

@@ -14,6 +14,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Run Tests from within the container
 - Add model dict output indexing in graph
 - Make lance upsert for added vectors
+- Make vectors normalized in inmemory vector database for cosine measure.
 
 #### New Features & Functionality
 - Add nightly image for pre-release testing in the cloud environment

superduperdb/vector_search/base.py

@@ -176,10 +176,9 @@ def cosine(x, y):
     '''
     Cosine similarity function for vector search
     '''
-    x = x.astype(float)
-    y = y.astype(float)
     x = x / numpy.linalg.norm(x, axis=1)[:, None]
-    y = y / numpy.linalg.norm(y, axis=1)[:, None]
+    # y which implies all vectors in vectordatabase
+    # has normalized vectors.
     return dot(x, y)
 
 

superduperdb/vector_search/in_memory.py

@@ -31,6 +31,10 @@ def __init__(
         self._cache: t.Sequence[VectorItem] = []
         self._CACHE_SIZE = 10000
 
+        self.measure = measure
+        if isinstance(measure, str):
+            self.measure = measures[measure]
+
         if h is not None:
             assert index is not None
             self._setup(h, index)
@@ -39,17 +43,19 @@ def __init__(
             self.index = None
             self.lookup = None
 
-        self.measure = measure
-        if isinstance(measure, str):
-            self.measure = measures[measure]
-
         self.identifier = identifier
 
     def __len__(self):
         return self.h.shape[0]
 
     def _setup(self, h, index):
-        self.h = numpy.array(h) if not isinstance(h, numpy.ndarray) else h
+        h = numpy.array(h) if not isinstance(h, numpy.ndarray) else h
+
+        if self.measure == 'cosine':
+            # Normalization is required for cosine, hence preparing
+            # all vectors in advance.
+            h = h / numpy.linalg.norm(h, axis=1)[:, None]
+        self.h = h
         self.index = index
         self.lookup = dict(zip(index, range(len(index))))