Fixed bug for mixed encoding types from numpy (#253)

paperqa/llms.py

@@ -126,7 +126,8 @@ async def embed_documents(self, client, texts) -> list[list[float]]:  # noqa: AR
         enc_batch = self.enc.encode_ordinary_batch(texts)
         # now get frequency of each token rel to length
         return [
-            np.bincount([xi % self.ndim for xi in x], minlength=self.ndim) / len(x)
+            np.bincount([xi % self.ndim for xi in x], minlength=self.ndim).astype(float)
+            / len(x)
             for x in enc_batch
         ]
 

paperqa/version.py

@@ -1 +1 @@
-__version__ = "4.2.0"
+__version__ = "4.2.1"

tests/test_paperqa.py

@@ -743,6 +743,9 @@ def test_sparse_embedding():
     )
     assert any(docs.docs["test"].embedding)  # type: ignore[arg-type]
 
+    # check the embeddings are the same size
+    assert np.shape(docs.texts[0].embedding) == np.shape(docs.texts[1].embedding)
+
     # test alias
     docs = Docs(embedding="sparse")
     assert docs._embedding_client is None
@@ -775,6 +778,9 @@ def test_hybrid_embedding():
     )
     assert any(docs.docs["test"].embedding)  # type: ignore[arg-type]
 
+    # check the embeddings are the same size
+    assert np.shape(docs.texts[0].embedding) == np.shape(docs.texts[1].embedding)
+
     # now try via alias
     docs = Docs(
         embedding="hybrid-text-embedding-3-small",