Merge pull request #1827 from weaviate/gen_openai_params

Add reasoning and verbosity to generative openai

Author: dirkkul

test/collection/test_classes_generative.py

@@ -322,6 +322,8 @@ def test_generative_parameters_images_parsing(
                 stop=["\n"],
                 temperature=0.5,
                 top_p=50,
+                reasoning_effort="high",
+                verbosity="low",
             )._to_grpc(
                 _GenerativeConfigRuntimeOptions(
                     return_metadata=True, images=[LOGO_ENCODED], image_properties=["image"]
@@ -344,6 +346,8 @@ def test_generative_parameters_images_parsing(
                     images=base_pb2.TextArray(values=[LOGO_ENCODED]),
                     image_properties=base_pb2.TextArray(values=["image"]),
                     is_azure=False,
+                    reasoning_effort=generative_pb2.GenerativeOpenAI.ReasoningEffort.REASONING_EFFORT_HIGH,
+                    verbosity=generative_pb2.GenerativeOpenAI.Verbosity.VERBOSITY_LOW,
                 ),
             ),
         ),

test/collection/test_config.py

@@ -838,6 +838,8 @@ def test_config_with_vectorizer_and_properties(
             temperature=0.5,
             top_p=0.5,
             base_url="https://api.openai.com",
+            reasoning_effort="high",
+            verbosity="verbose",
         ),
         {
             "generative-openai": {
@@ -848,6 +850,8 @@ def test_config_with_vectorizer_and_properties(
                 "temperatureProperty": 0.5,
                 "topPProperty": 0.5,
                 "baseURL": "https://api.openai.com/",
+                "reasoningEffort": "high",
+                "verbosity": "verbose",
             }
         },
     ),

weaviate/collections/classes/config.py

@@ -82,6 +82,19 @@
     "sagemaker",
 ]
 
+OpenAiVerbosity: TypeAlias = Literal[
+    "low",
+    "medium",
+    "high",
+]
+
+OpenAiReasoningEffort: TypeAlias = Literal[
+    "minimal",
+    "low",
+    "medium",
+    "high",
+]
+
 
 class ConsistencyLevel(str, BaseEnum):
     """The consistency levels when writing to Weaviate with replication enabled.
@@ -416,6 +429,8 @@ def _to_dict(self) -> Dict[str, Any]:
 
 class _GenerativeOpenAIConfig(_GenerativeOpenAIConfigBase):
     model: Optional[str]
+    verbosity: Optional[str]
+    reasoningEffort: Optional[str]
 
 
 class _GenerativeAzureOpenAIConfig(_GenerativeOpenAIConfigBase):
@@ -711,6 +726,9 @@ def openai(
         temperature: Optional[float] = None,
         top_p: Optional[float] = None,
         base_url: Optional[AnyHttpUrl] = None,
+        *,
+        verbosity: Optional[Union[OpenAiVerbosity, str]] = None,
+        reasoning_effort: Optional[Union[OpenAiReasoningEffort, str]] = None,
     ) -> _GenerativeProvider:
         """Create a `_GenerativeOpenAIConfig` object for use when performing AI generation using the `generative-openai` module.
 
@@ -725,6 +743,8 @@ def openai(
             temperature: The temperature to use. Defaults to `None`, which uses the server-defined default
             top_p: The top P to use. Defaults to `None`, which uses the server-defined default
             base_url: The base URL where the API request should go. Defaults to `None`, which uses the server-defined default
+            verbosity: The verbosity level to use. One of 'low', 'medium', 'high'. Defaults to `None`, which uses the server-defined default
+            reasoning_effort: The reasoning effort level to use. One of 'minimal', 'low
         """
         return _GenerativeOpenAIConfig(
             baseURL=base_url,
@@ -734,6 +754,8 @@ def openai(
             presencePenaltyProperty=presence_penalty,
             temperatureProperty=temperature,
             topPProperty=top_p,
+            verbosity=verbosity,
+            reasoningEffort=reasoning_effort,
         )
 
     @staticmethod

weaviate/collections/classes/generative.py

@@ -9,6 +9,8 @@
 from weaviate.collections.classes.config import (
     AWSService,
     GenerativeSearches,
+    OpenAiReasoningEffort,
+    OpenAiVerbosity,
     _EnumLikeStr,
 )
 from weaviate.exceptions import WeaviateInvalidInputError
@@ -319,6 +321,8 @@ class _GenerativeOpenAI(_GenerativeConfigRuntime):
     stop: Optional[List[str]]
     temperature: Optional[float]
     top_p: Optional[float]
+    verbosity: Optional[Union[OpenAiVerbosity, str]]
+    reasoning_effort: Optional[Union[OpenAiReasoningEffort, str]]
 
     def _to_grpc(self, opts: _GenerativeConfigRuntimeOptions) -> generative_pb2.GenerativeProvider:
         return generative_pb2.GenerativeProvider(
@@ -338,9 +342,37 @@ def _to_grpc(self, opts: _GenerativeConfigRuntimeOptions) -> generative_pb2.Gene
                 is_azure=self.is_azure,
                 images=_to_text_array(opts.images),
                 image_properties=_to_text_array(opts.image_properties),
+                verbosity=self.__verbosity(),
+                reasoning_effort=self.__reasoning_effort(),
             ),
         )
 
+    def __verbosity(self):
+        if self.verbosity is None:
+            return None
+
+        if self.verbosity == "low":
+            return generative_pb2.GenerativeOpenAI.Verbosity.VERBOSITY_LOW
+        if self.verbosity == "medium":
+            return generative_pb2.GenerativeOpenAI.Verbosity.VERBOSITY_MEDIUM
+        if self.verbosity == "high":
+            return generative_pb2.GenerativeOpenAI.Verbosity.VERBOSITY_HIGH
+        raise WeaviateInvalidInputError(f"Invalid verbosity value: {self.verbosity}")
+
+    def __reasoning_effort(self):
+        if self.reasoning_effort is None:
+            return None
+
+        if self.reasoning_effort == "minimal":
+            return generative_pb2.GenerativeOpenAI.ReasoningEffort.REASONING_EFFORT_MINIMAL
+        if self.reasoning_effort == "low":
+            return generative_pb2.GenerativeOpenAI.ReasoningEffort.REASONING_EFFORT_LOW
+        if self.reasoning_effort == "medium":
+            return generative_pb2.GenerativeOpenAI.ReasoningEffort.REASONING_EFFORT_MEDIUM
+        if self.reasoning_effort == "high":
+            return generative_pb2.GenerativeOpenAI.ReasoningEffort.REASONING_EFFORT_HIGH
+        raise WeaviateInvalidInputError(f"Invalid reasoning_effort value: {self.reasoning_effort}")
+
 
 class _GenerativeGoogle(_GenerativeConfigRuntime):
     generative: Union[GenerativeSearches, _EnumLikeStr] = Field(
@@ -763,10 +795,12 @@ def openai(
         max_tokens: Optional[int] = None,
         model: Optional[str] = None,
         presence_penalty: Optional[float] = None,
+        reasoning_effort: Optional[Union[OpenAiReasoningEffort, str]] = None,
         resource_name: Optional[str] = None,
         stop: Optional[List[str]] = None,
         temperature: Optional[float] = None,
         top_p: Optional[float] = None,
+        verbosity: Optional[Union[OpenAiVerbosity, str]] = None,
     ) -> _GenerativeConfigRuntime:
         """Create a `_GenerativeOpenAI` object for use when performing AI generation using the OpenAI-backed `generative-openai` module.
 
@@ -781,10 +815,12 @@ def openai(
             max_tokens: The maximum number of tokens to generate. Defaults to `None`, which uses the server-defined default
             model: The model to use. Defaults to `None`, which uses the server-defined default
             presence_penalty: The presence penalty to use. Defaults to `None`, which uses the server-defined default
+            reasoning_effort: The reasoning effort to use. Defaults to `None`, which uses the server-defined default
             resource_name: The name of the OpenAI resource to use. Defaults to `None`, which uses the server-defined default
             stop: The stop sequences to use. Defaults to `None`, which uses the server-defined default
             temperature: The temperature to use. Defaults to `None`, which uses the server-defined default
             top_p: The top P to use. Defaults to `None`, which uses the server-defined default
+            verbosity: The verbosity to use. Defaults to `None`, which uses the server-defined default
         """
         return _GenerativeOpenAI(
             api_version=api_version,
@@ -799,6 +835,8 @@ def openai(
             temperature=temperature,
             top_p=top_p,
             is_azure=False,
+            verbosity=verbosity,
+            reasoning_effort=reasoning_effort,
         )
 
     @staticmethod
@@ -847,6 +885,8 @@ def azure_openai(
             temperature=temperature,
             top_p=top_p,
             is_azure=True,
+            verbosity=None,
+            reasoning_effort=None,
         )
 
     @staticmethod

weaviate/proto/v1/v4216/v1/generative_pb2.py

@@ -1,5 +1,6 @@
 from weaviate.proto.v1.v4216.v1 import base_pb2 as _base_pb2
 from google.protobuf.internal import containers as _containers
+from google.protobuf.internal import enum_type_wrapper as _enum_type_wrapper
 from google.protobuf import descriptor as _descriptor
 from google.protobuf import message as _message
 from typing import ClassVar as _ClassVar, Iterable as _Iterable, Mapping as _Mapping, Optional as _Optional, Union as _Union
@@ -183,7 +184,29 @@ class GenerativeOllama(_message.Message):
     def __init__(self, api_endpoint: _Optional[str] = ..., model: _Optional[str] = ..., temperature: _Optional[float] = ..., images: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., image_properties: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ...) -> None: ...
 
 class GenerativeOpenAI(_message.Message):
-    __slots__ = ["frequency_penalty", "max_tokens", "model", "n", "presence_penalty", "stop", "temperature", "top_p", "base_url", "api_version", "resource_name", "deployment_id", "is_azure", "images", "image_properties"]
+    __slots__ = ["frequency_penalty", "max_tokens", "model", "n", "presence_penalty", "stop", "temperature", "top_p", "base_url", "api_version", "resource_name", "deployment_id", "is_azure", "images", "image_properties", "reasoning_effort", "verbosity"]
+    class ReasoningEffort(int, metaclass=_enum_type_wrapper.EnumTypeWrapper):
+        __slots__ = []
+        REASONING_EFFORT_UNSPECIFIED: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_MINIMAL: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_LOW: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_MEDIUM: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_HIGH: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+    REASONING_EFFORT_UNSPECIFIED: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_MINIMAL: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_LOW: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_MEDIUM: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_HIGH: GenerativeOpenAI.ReasoningEffort
+    class Verbosity(int, metaclass=_enum_type_wrapper.EnumTypeWrapper):
+        __slots__ = []
+        VERBOSITY_UNSPECIFIED: _ClassVar[GenerativeOpenAI.Verbosity]
+        VERBOSITY_LOW: _ClassVar[GenerativeOpenAI.Verbosity]
+        VERBOSITY_MEDIUM: _ClassVar[GenerativeOpenAI.Verbosity]
+        VERBOSITY_HIGH: _ClassVar[GenerativeOpenAI.Verbosity]
+    VERBOSITY_UNSPECIFIED: GenerativeOpenAI.Verbosity
+    VERBOSITY_LOW: GenerativeOpenAI.Verbosity
+    VERBOSITY_MEDIUM: GenerativeOpenAI.Verbosity
+    VERBOSITY_HIGH: GenerativeOpenAI.Verbosity
     FREQUENCY_PENALTY_FIELD_NUMBER: _ClassVar[int]
     MAX_TOKENS_FIELD_NUMBER: _ClassVar[int]
     MODEL_FIELD_NUMBER: _ClassVar[int]
@@ -199,6 +222,8 @@ class GenerativeOpenAI(_message.Message):
     IS_AZURE_FIELD_NUMBER: _ClassVar[int]
     IMAGES_FIELD_NUMBER: _ClassVar[int]
     IMAGE_PROPERTIES_FIELD_NUMBER: _ClassVar[int]
+    REASONING_EFFORT_FIELD_NUMBER: _ClassVar[int]
+    VERBOSITY_FIELD_NUMBER: _ClassVar[int]
     frequency_penalty: float
     max_tokens: int
     model: str
@@ -214,7 +239,9 @@ class GenerativeOpenAI(_message.Message):
     is_azure: bool
     images: _base_pb2.TextArray
     image_properties: _base_pb2.TextArray
-    def __init__(self, frequency_penalty: _Optional[float] = ..., max_tokens: _Optional[int] = ..., model: _Optional[str] = ..., n: _Optional[int] = ..., presence_penalty: _Optional[float] = ..., stop: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., temperature: _Optional[float] = ..., top_p: _Optional[float] = ..., base_url: _Optional[str] = ..., api_version: _Optional[str] = ..., resource_name: _Optional[str] = ..., deployment_id: _Optional[str] = ..., is_azure: bool = ..., images: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., image_properties: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ...) -> None: ...
+    reasoning_effort: GenerativeOpenAI.ReasoningEffort
+    verbosity: GenerativeOpenAI.Verbosity
+    def __init__(self, frequency_penalty: _Optional[float] = ..., max_tokens: _Optional[int] = ..., model: _Optional[str] = ..., n: _Optional[int] = ..., presence_penalty: _Optional[float] = ..., stop: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., temperature: _Optional[float] = ..., top_p: _Optional[float] = ..., base_url: _Optional[str] = ..., api_version: _Optional[str] = ..., resource_name: _Optional[str] = ..., deployment_id: _Optional[str] = ..., is_azure: bool = ..., images: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., image_properties: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., reasoning_effort: _Optional[_Union[GenerativeOpenAI.ReasoningEffort, str]] = ..., verbosity: _Optional[_Union[GenerativeOpenAI.Verbosity, str]] = ...) -> None: ...
 
 class GenerativeGoogle(_message.Message):
     __slots__ = ["frequency_penalty", "max_tokens", "model", "presence_penalty", "temperature", "top_k", "top_p", "stop_sequences", "api_endpoint", "project_id", "endpoint_id", "region", "images", "image_properties"]

weaviate/proto/v1/v4216/v1/generative_pb2.pyi

@@ -1,5 +1,6 @@
 from weaviate.proto.v1.v5261.v1 import base_pb2 as _base_pb2
 from google.protobuf.internal import containers as _containers
+from google.protobuf.internal import enum_type_wrapper as _enum_type_wrapper
 from google.protobuf import descriptor as _descriptor
 from google.protobuf import message as _message
 from typing import ClassVar as _ClassVar, Iterable as _Iterable, Mapping as _Mapping, Optional as _Optional, Union as _Union
@@ -183,7 +184,29 @@ class GenerativeOllama(_message.Message):
     def __init__(self, api_endpoint: _Optional[str] = ..., model: _Optional[str] = ..., temperature: _Optional[float] = ..., images: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., image_properties: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ...) -> None: ...
 
 class GenerativeOpenAI(_message.Message):
-    __slots__ = ("frequency_penalty", "max_tokens", "model", "n", "presence_penalty", "stop", "temperature", "top_p", "base_url", "api_version", "resource_name", "deployment_id", "is_azure", "images", "image_properties")
+    __slots__ = ("frequency_penalty", "max_tokens", "model", "n", "presence_penalty", "stop", "temperature", "top_p", "base_url", "api_version", "resource_name", "deployment_id", "is_azure", "images", "image_properties", "reasoning_effort", "verbosity")
+    class ReasoningEffort(int, metaclass=_enum_type_wrapper.EnumTypeWrapper):
+        __slots__ = ()
+        REASONING_EFFORT_UNSPECIFIED: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_MINIMAL: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_LOW: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_MEDIUM: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+        REASONING_EFFORT_HIGH: _ClassVar[GenerativeOpenAI.ReasoningEffort]
+    REASONING_EFFORT_UNSPECIFIED: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_MINIMAL: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_LOW: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_MEDIUM: GenerativeOpenAI.ReasoningEffort
+    REASONING_EFFORT_HIGH: GenerativeOpenAI.ReasoningEffort
+    class Verbosity(int, metaclass=_enum_type_wrapper.EnumTypeWrapper):
+        __slots__ = ()
+        VERBOSITY_UNSPECIFIED: _ClassVar[GenerativeOpenAI.Verbosity]
+        VERBOSITY_LOW: _ClassVar[GenerativeOpenAI.Verbosity]
+        VERBOSITY_MEDIUM: _ClassVar[GenerativeOpenAI.Verbosity]
+        VERBOSITY_HIGH: _ClassVar[GenerativeOpenAI.Verbosity]
+    VERBOSITY_UNSPECIFIED: GenerativeOpenAI.Verbosity
+    VERBOSITY_LOW: GenerativeOpenAI.Verbosity
+    VERBOSITY_MEDIUM: GenerativeOpenAI.Verbosity
+    VERBOSITY_HIGH: GenerativeOpenAI.Verbosity
     FREQUENCY_PENALTY_FIELD_NUMBER: _ClassVar[int]
     MAX_TOKENS_FIELD_NUMBER: _ClassVar[int]
     MODEL_FIELD_NUMBER: _ClassVar[int]
@@ -199,6 +222,8 @@ class GenerativeOpenAI(_message.Message):
     IS_AZURE_FIELD_NUMBER: _ClassVar[int]
     IMAGES_FIELD_NUMBER: _ClassVar[int]
     IMAGE_PROPERTIES_FIELD_NUMBER: _ClassVar[int]
+    REASONING_EFFORT_FIELD_NUMBER: _ClassVar[int]
+    VERBOSITY_FIELD_NUMBER: _ClassVar[int]
     frequency_penalty: float
     max_tokens: int
     model: str
@@ -214,7 +239,9 @@ class GenerativeOpenAI(_message.Message):
     is_azure: bool
     images: _base_pb2.TextArray
     image_properties: _base_pb2.TextArray
-    def __init__(self, frequency_penalty: _Optional[float] = ..., max_tokens: _Optional[int] = ..., model: _Optional[str] = ..., n: _Optional[int] = ..., presence_penalty: _Optional[float] = ..., stop: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., temperature: _Optional[float] = ..., top_p: _Optional[float] = ..., base_url: _Optional[str] = ..., api_version: _Optional[str] = ..., resource_name: _Optional[str] = ..., deployment_id: _Optional[str] = ..., is_azure: bool = ..., images: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., image_properties: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ...) -> None: ...
+    reasoning_effort: GenerativeOpenAI.ReasoningEffort
+    verbosity: GenerativeOpenAI.Verbosity
+    def __init__(self, frequency_penalty: _Optional[float] = ..., max_tokens: _Optional[int] = ..., model: _Optional[str] = ..., n: _Optional[int] = ..., presence_penalty: _Optional[float] = ..., stop: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., temperature: _Optional[float] = ..., top_p: _Optional[float] = ..., base_url: _Optional[str] = ..., api_version: _Optional[str] = ..., resource_name: _Optional[str] = ..., deployment_id: _Optional[str] = ..., is_azure: bool = ..., images: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., image_properties: _Optional[_Union[_base_pb2.TextArray, _Mapping]] = ..., reasoning_effort: _Optional[_Union[GenerativeOpenAI.ReasoningEffort, str]] = ..., verbosity: _Optional[_Union[GenerativeOpenAI.Verbosity, str]] = ...) -> None: ...
 
 class GenerativeGoogle(_message.Message):
     __slots__ = ("frequency_penalty", "max_tokens", "model", "presence_penalty", "temperature", "top_k", "top_p", "stop_sequences", "api_endpoint", "project_id", "endpoint_id", "region", "images", "image_properties")