new: preserve embeddings in a type set by their model (#492)

* new: preserve embeddings in a type set by their model

* fix: remove type coercion

* fix: remove redundant type

* fix: fix random data type in tests

fastembed/image/onnx_embedding.py

@@ -1,6 +1,5 @@
 from typing import Any, Iterable, Optional, Sequence, Type, Union
 
-import numpy as np
 
 from fastembed.common.types import NumpyArray
 from fastembed.common import ImageInput, OnnxProvider
@@ -195,7 +194,7 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
         return onnx_input
 
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
-        return normalize(output.model_output).astype(np.float32)
+        return normalize(output.model_output)
 
 
 class OnnxImageEmbeddingWorker(ImageEmbeddingWorker[NumpyArray]):

fastembed/late_interaction/colbert.py

@@ -46,7 +46,7 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         self, output: OnnxOutputContext, is_doc: bool = True
     ) -> Iterable[NumpyArray]:
         if not is_doc:
-            return output.model_output.astype(np.float32)
+            return output.model_output
 
         if output.input_ids is None or output.attention_mask is None:
             raise ValueError(
@@ -58,11 +58,11 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
                 if token_id in self.skip_list or token_id == self.pad_token_id:
                     output.attention_mask[i, j] = 0
 
-        output.model_output *= np.expand_dims(output.attention_mask, 2).astype(np.float32)
+        output.model_output *= np.expand_dims(output.attention_mask, 2)
         norm = np.linalg.norm(output.model_output, ord=2, axis=2, keepdims=True)
         norm_clamped = np.maximum(norm, 1e-12)
         output.model_output /= norm_clamped
-        return output.model_output.astype(np.float32)
+        return output.model_output
 
     def _preprocess_onnx_input(
         self, onnx_input: dict[str, NumpyArray], is_doc: bool = True, **kwargs: Any

fastembed/late_interaction_multimodal/colpali.py

@@ -142,7 +142,7 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         assert self.model_description.dim is not None, "Model dim is not defined"
         return output.model_output.reshape(
             output.model_output.shape[0], -1, self.model_description.dim
-        ).astype(np.float32)
+        )
 
     def _post_process_onnx_text_output(
         self,
@@ -157,7 +157,7 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         Returns:
             Iterable[NumpyArray]: Post-processed output as NumPy arrays.
         """
-        return output.model_output.astype(np.float32)
+        return output.model_output
 
     def tokenize(self, documents: list[str], **kwargs: Any) -> list[Encoding]:
         texts_query: list[str] = []

fastembed/text/onnx_embedding.py

@@ -1,6 +1,5 @@
 from typing import Any, Iterable, Optional, Sequence, Type, Union
 
-import numpy as np
 from fastembed.common.types import NumpyArray, OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.utils import define_cache_dir, normalize
@@ -313,7 +312,7 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
             processed_embeddings = embeddings
         else:
             raise ValueError(f"Unsupported embedding shape: {embeddings.shape}")
-        return normalize(processed_embeddings).astype(np.float32)
+        return normalize(processed_embeddings)
 
     def load_onnx_model(self) -> None:
         self._load_onnx_model(

fastembed/text/pooled_embedding.py

@@ -116,7 +116,7 @@ class PooledEmbedding(OnnxTextEmbedding):
 
         embeddings = output.model_output
         attn_mask = output.attention_mask
-        return self.mean_pooling(embeddings, attn_mask).astype(np.float32)
+        return self.mean_pooling(embeddings, attn_mask)
 
 
 class PooledEmbeddingWorker(OnnxTextEmbeddingWorker):

fastembed/text/pooled_normalized_embedding.py

@@ -1,6 +1,5 @@
 from typing import Any, Iterable, Type
 
-import numpy as np
 
 from fastembed.common.types import NumpyArray
 from fastembed.common.onnx_model import OnnxOutputContext
@@ -145,7 +144,7 @@ class PooledNormalizedEmbedding(PooledEmbedding):
 
         embeddings = output.model_output
         attn_mask = output.attention_mask
-        return normalize(self.mean_pooling(embeddings, attn_mask)).astype(np.float32)
+        return normalize(self.mean_pooling(embeddings, attn_mask))
 
 
 class PooledNormalizedEmbeddingWorker(OnnxTextEmbeddingWorker):

tests/test_custom_models.py

@@ -91,15 +91,11 @@ def test_mock_add_custom_models():
     expected_output = {
         f"{PoolingType.MEAN.lower()}-normalized": normalize(
             mean_pooling(dummy_token_embedding, dummy_attention_mask)
-        ).astype(np.float32),
+        ),
         f"{PoolingType.MEAN.lower()}": mean_pooling(dummy_token_embedding, dummy_attention_mask),
-        f"{PoolingType.CLS.lower()}-normalized": normalize(dummy_token_embedding[:, 0]).astype(
-            np.float32
-        ),
+        f"{PoolingType.CLS.lower()}-normalized": normalize(dummy_token_embedding[:, 0]),
         f"{PoolingType.CLS.lower()}": dummy_token_embedding[:, 0],
-        f"{PoolingType.DISABLED.lower()}-normalized": normalize(dummy_pooled_embedding).astype(
-            np.float32
-        ),
+        f"{PoolingType.DISABLED.lower()}-normalized": normalize(dummy_pooled_embedding),
         f"{PoolingType.DISABLED.lower()}": dummy_pooled_embedding,
     }