feat: replace the output_format with a boolean flag
Browse files- modeling_jina_embeddings_v4.py +11 -26
modeling_jina_embeddings_v4.py
CHANGED
|
@@ -30,11 +30,6 @@ class PromptType(str, Enum):
|
|
| 30 |
passage = "passage"
|
| 31 |
|
| 32 |
|
| 33 |
-
class VectorOutputFormat(str, Enum):
|
| 34 |
-
SINGLE = "single"
|
| 35 |
-
MULTIPLE = "multiple"
|
| 36 |
-
|
| 37 |
-
|
| 38 |
PREFIX_DICT = {"query": "Query", "passage": "Passage"}
|
| 39 |
|
| 40 |
|
|
@@ -325,7 +320,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 325 |
task_label: Union[str, List[str]],
|
| 326 |
processor_fn: Callable,
|
| 327 |
desc: str,
|
| 328 |
-
|
| 329 |
return_numpy: bool = False,
|
| 330 |
batch_size: int = 32,
|
| 331 |
truncate_dim: Optional[int] = None,
|
|
@@ -345,8 +340,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 345 |
device_type=torch.device(self.device).type, dtype=torch.bfloat16
|
| 346 |
):
|
| 347 |
embeddings = self(**batch, task_label=task_label)
|
| 348 |
-
|
| 349 |
-
if output_format_str == VectorOutputFormat.SINGLE.value:
|
| 350 |
embeddings = embeddings.single_vec_emb
|
| 351 |
if truncate_dim is not None:
|
| 352 |
embeddings = embeddings[:, :truncate_dim]
|
|
@@ -363,7 +357,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 363 |
|
| 364 |
def _validate_encoding_params(
|
| 365 |
self,
|
| 366 |
-
|
| 367 |
truncate_dim: Optional[int] = None,
|
| 368 |
prompt_name: Optional[str] = None,
|
| 369 |
) -> Dict[str, Any]:
|
|
@@ -380,17 +374,8 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 380 |
else PREFIX_DICT["query"]
|
| 381 |
)
|
| 382 |
|
| 383 |
-
|
| 384 |
-
|
| 385 |
-
encode_kwargs["output_format"] = output_format.value
|
| 386 |
-
else:
|
| 387 |
-
try:
|
| 388 |
-
output_format_enum = VectorOutputFormat(output_format)
|
| 389 |
-
encode_kwargs["output_format"] = output_format_enum.value
|
| 390 |
-
except ValueError:
|
| 391 |
-
raise ValueError(
|
| 392 |
-
f"Invalid output_format: {output_format}. Must be one of {[v.value for v in VectorOutputFormat]}."
|
| 393 |
-
)
|
| 394 |
|
| 395 |
truncate_dim = truncate_dim or self.config.truncate_dim
|
| 396 |
if truncate_dim is not None and truncate_dim not in self.config.matryoshka_dims:
|
|
@@ -423,7 +408,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 423 |
task: Optional[str] = None,
|
| 424 |
max_length: int = 8192,
|
| 425 |
batch_size: int = 8,
|
| 426 |
-
|
| 427 |
return_numpy: bool = False,
|
| 428 |
truncate_dim: Optional[int] = None,
|
| 429 |
prompt_name: Optional[str] = None,
|
|
@@ -435,7 +420,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 435 |
texts: text or list of text strings to encode
|
| 436 |
max_length: Maximum token length for text processing
|
| 437 |
batch_size: Number of texts to process at once
|
| 438 |
-
|
| 439 |
return_numpy: Whether to return numpy arrays instead of torch tensors
|
| 440 |
truncate_dim: Dimension to truncate embeddings to (128, 256, 512, or 1024)
|
| 441 |
prompt_name: Type of text being encoded ('query' or 'passage')
|
|
@@ -445,7 +430,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 445 |
"""
|
| 446 |
prompt_name = prompt_name or "query"
|
| 447 |
encode_kwargs = self._validate_encoding_params(
|
| 448 |
-
|
| 449 |
)
|
| 450 |
|
| 451 |
task = self._validate_task(task)
|
|
@@ -490,7 +475,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 490 |
images: Union[str, Image.Image, List[Union[str, Image.Image]]],
|
| 491 |
task: Optional[str] = None,
|
| 492 |
batch_size: int = 8,
|
| 493 |
-
|
| 494 |
return_numpy: bool = False,
|
| 495 |
truncate_dim: Optional[int] = None,
|
| 496 |
max_pixels: Optional[int] = None,
|
|
@@ -501,7 +486,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 501 |
Args:
|
| 502 |
images: image(s) to encode, can be PIL Image(s), URL(s), or local file path(s)
|
| 503 |
batch_size: Number of images to process at once
|
| 504 |
-
|
| 505 |
return_numpy: Whether to return numpy arrays instead of torch tensors
|
| 506 |
truncate_dim: Dimension to truncate embeddings to (128, 256, 512, or 1024)
|
| 507 |
max_pixels: Maximum number of pixels to process per image
|
|
@@ -514,7 +499,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
| 514 |
self.processor.image_processor.max_pixels = (
|
| 515 |
max_pixels # change during encoding
|
| 516 |
)
|
| 517 |
-
encode_kwargs = self._validate_encoding_params(
|
| 518 |
task = self._validate_task(task)
|
| 519 |
|
| 520 |
# Convert single image to list
|
|
|
|
| 30 |
passage = "passage"
|
| 31 |
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
PREFIX_DICT = {"query": "Query", "passage": "Passage"}
|
| 34 |
|
| 35 |
|
|
|
|
| 320 |
task_label: Union[str, List[str]],
|
| 321 |
processor_fn: Callable,
|
| 322 |
desc: str,
|
| 323 |
+
return_multivector: bool = False,
|
| 324 |
return_numpy: bool = False,
|
| 325 |
batch_size: int = 32,
|
| 326 |
truncate_dim: Optional[int] = None,
|
|
|
|
| 340 |
device_type=torch.device(self.device).type, dtype=torch.bfloat16
|
| 341 |
):
|
| 342 |
embeddings = self(**batch, task_label=task_label)
|
| 343 |
+
if not return_multivector:
|
|
|
|
| 344 |
embeddings = embeddings.single_vec_emb
|
| 345 |
if truncate_dim is not None:
|
| 346 |
embeddings = embeddings[:, :truncate_dim]
|
|
|
|
| 357 |
|
| 358 |
def _validate_encoding_params(
|
| 359 |
self,
|
| 360 |
+
return_multivector: Optional[bool] = None,
|
| 361 |
truncate_dim: Optional[int] = None,
|
| 362 |
prompt_name: Optional[str] = None,
|
| 363 |
) -> Dict[str, Any]:
|
|
|
|
| 374 |
else PREFIX_DICT["query"]
|
| 375 |
)
|
| 376 |
|
| 377 |
+
return_multivector = return_multivector or False
|
| 378 |
+
encode_kwargs["return_multivector"] = return_multivector
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 379 |
|
| 380 |
truncate_dim = truncate_dim or self.config.truncate_dim
|
| 381 |
if truncate_dim is not None and truncate_dim not in self.config.matryoshka_dims:
|
|
|
|
| 408 |
task: Optional[str] = None,
|
| 409 |
max_length: int = 8192,
|
| 410 |
batch_size: int = 8,
|
| 411 |
+
return_multivector: bool = False,
|
| 412 |
return_numpy: bool = False,
|
| 413 |
truncate_dim: Optional[int] = None,
|
| 414 |
prompt_name: Optional[str] = None,
|
|
|
|
| 420 |
texts: text or list of text strings to encode
|
| 421 |
max_length: Maximum token length for text processing
|
| 422 |
batch_size: Number of texts to process at once
|
| 423 |
+
return_multivector: Whether to return multi-vector embeddings instead of single-vector embeddings
|
| 424 |
return_numpy: Whether to return numpy arrays instead of torch tensors
|
| 425 |
truncate_dim: Dimension to truncate embeddings to (128, 256, 512, or 1024)
|
| 426 |
prompt_name: Type of text being encoded ('query' or 'passage')
|
|
|
|
| 430 |
"""
|
| 431 |
prompt_name = prompt_name or "query"
|
| 432 |
encode_kwargs = self._validate_encoding_params(
|
| 433 |
+
return_multivector=return_multivector, truncate_dim=truncate_dim, prompt_name=prompt_name
|
| 434 |
)
|
| 435 |
|
| 436 |
task = self._validate_task(task)
|
|
|
|
| 475 |
images: Union[str, Image.Image, List[Union[str, Image.Image]]],
|
| 476 |
task: Optional[str] = None,
|
| 477 |
batch_size: int = 8,
|
| 478 |
+
return_multivector: bool = False,
|
| 479 |
return_numpy: bool = False,
|
| 480 |
truncate_dim: Optional[int] = None,
|
| 481 |
max_pixels: Optional[int] = None,
|
|
|
|
| 486 |
Args:
|
| 487 |
images: image(s) to encode, can be PIL Image(s), URL(s), or local file path(s)
|
| 488 |
batch_size: Number of images to process at once
|
| 489 |
+
return_multivector: Whether to return multi-vector embeddings instead of single-vector embeddings
|
| 490 |
return_numpy: Whether to return numpy arrays instead of torch tensors
|
| 491 |
truncate_dim: Dimension to truncate embeddings to (128, 256, 512, or 1024)
|
| 492 |
max_pixels: Maximum number of pixels to process per image
|
|
|
|
| 499 |
self.processor.image_processor.max_pixels = (
|
| 500 |
max_pixels # change during encoding
|
| 501 |
)
|
| 502 |
+
encode_kwargs = self._validate_encoding_params(return_multivector=return_multivector, truncate_dim=truncate_dim)
|
| 503 |
task = self._validate_task(task)
|
| 504 |
|
| 505 |
# Convert single image to list
|