From 3e0ef1cc8a6e6cef44056fd18f4e94089fc90311 Mon Sep 17 00:00:00 2001
From: bogdankostic <bogdankostic@web.de>
Date: Mon, 3 Jan 2022 17:14:51 +0100
Subject: [PATCH] Fix Numba TypingError in `normalize_embedding` for cosine
 similarity (#1933)

* Fix Numba TypingError

* Add latest docstring and tutorial changes

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
---
 docs/_src/api/api/document_store.md |  4 +++-
 haystack/document_stores/base.py    | 15 ++++++++++-----
 2 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/docs/_src/api/api/document_store.md b/docs/_src/api/api/document_store.md
index 9d0009c0b..8464bc47b 100644
--- a/docs/_src/api/api/document_store.md
+++ b/docs/_src/api/api/document_store.md
@@ -130,11 +130,13 @@ object, provided that they have the same product_id (to be found in Label.meta["
 #### normalize\_embedding
 
 ```python
+ | @staticmethod
  | @njit
  | normalize_embedding(emb: np.ndarray) -> None
 ```
 
-Performs L2 normalization of embeddings vector inplace. Input can be a single vector (1D array) or a matrix (2D array).
+Performs L2 normalization of embeddings vector inplace. Input can be a single vector (1D array) or a matrix
+(2D array).
 
 <a name="base.BaseDocumentStore.add_eval_data"></a>
 #### add\_eval\_data
diff --git a/haystack/document_stores/base.py b/haystack/document_stores/base.py
index 9526abe5a..93add2520 100644
--- a/haystack/document_stores/base.py
+++ b/haystack/document_stores/base.py
@@ -232,22 +232,27 @@ class BaseDocumentStore(BaseComponent):
         headers: Optional[Dict[str, str]] = None) -> int:
         pass
 
+    @staticmethod
     @njit#(fastmath=True)
-    def normalize_embedding(self, emb: np.ndarray) -> None:
+    def normalize_embedding(emb: np.ndarray) -> None:
         """
-        Performs L2 normalization of embeddings vector inplace. Input can be a single vector (1D array) or a matrix (2D array).
+        Performs L2 normalization of embeddings vector inplace. Input can be a single vector (1D array) or a matrix
+        (2D array).
         """
         # Might be extended to other normalizations in future
 
         # Single vec
         if len(emb.shape) == 1:
-            norm = np.sqrt(emb.dot(emb)) #faster than np.linalg.norm()
+            norm = np.sqrt(emb.dot(emb))  # faster than np.linalg.norm()
             if norm != 0.0:
                 emb /= norm
         # 2D matrix
         else:
-            norm = np.linalg.norm(emb, axis=1)
-            emb /= norm[:, None]
+            for vec in emb:
+                vec = np.ascontiguousarray(vec)
+                norm = np.sqrt(vec.dot(vec))
+                if norm != 0.0:
+                    vec /= norm
 
     def finalize_raw_score(self, raw_score: float, similarity: Optional[str]) -> float:
         if similarity == "cosine":