feat: add GPT-5 model family support #92 (#140)

Co-authored-by: Bahdan Kapionkin <bahdan_kapionkin@epam.com>

Author: bahdan111

statgpt/common/config/__init__.py

@@ -1,4 +1,4 @@
-from .llm_models import EmbeddingModelsEnum, LLMModelsEnum
+from .llm_models import EmbeddingModelsEnum, LLMModelsEnum, ReasoningEffortEnum, VerbosityEnum
 from .logging import LoggingConfig, logger, multiline_logger
 from .versions import Versions
 

statgpt/common/config/llm_models.py

@@ -6,6 +6,26 @@ class EmbeddingModelsEnum(StrEnum):
     TEXT_EMBEDDING_3_LARGE = "text-embedding-3-large"
 
 
+class ReasoningEffortEnum(StrEnum):
+    """Reasoning effort levels for GPT-5 models."""
+
+    NONE = "none"
+    """No reasoning mode - standard inference."""
+    MINIMAL = "minimal"
+    LOW = "low"
+    MEDIUM = "medium"
+    HIGH = "high"
+    XHIGH = "xhigh"
+
+
+class VerbosityEnum(StrEnum):
+    """Output verbosity levels for GPT-5 models."""
+
+    LOW = "low"
+    MEDIUM = "medium"
+    HIGH = "high"
+
+
 class LLMModelsEnum(StrEnum):
     # Gemini models
     GEMINI_2_0_FLASH_LITE_001 = "gemini-2.0-flash-lite-001"
@@ -30,6 +50,11 @@ class LLMModelsEnum(StrEnum):
     GPT_4_1_MINI_2025_04_14 = "gpt-4.1-mini-2025-04-14"
     GPT_4_1_NANO_2025_04_14 = "gpt-4.1-nano-2025-04-14"
 
+    # GPT-5 models
+    GPT_5_MINI_2025_08_07 = "gpt-5-mini-2025-08-07"
+    GPT_5_1_2025_11_13 = "gpt-5.1-2025-11-13"
+    GPT_5_2_2025_12_11 = "gpt-5.2-2025-12-11"
+
     @property
     def deployment_id(self) -> str:
         return os.getenv(f"LLM_MODELS_{self.name}", self.value)
@@ -42,3 +67,12 @@ def is_gpt_41_family(self) -> bool:
             LLMModelsEnum.GPT_4_1_MINI_2025_04_14,
             LLMModelsEnum.GPT_4_1_NANO_2025_04_14,
         }
+
+    @property
+    def is_gpt_5_family(self) -> bool:
+        """Check if the model belongs to the GPT-5 family."""
+        return self in {
+            LLMModelsEnum.GPT_5_MINI_2025_08_07,
+            LLMModelsEnum.GPT_5_1_2025_11_13,
+            LLMModelsEnum.GPT_5_2_2025_12_11,
+        }

statgpt/common/models/models.py

@@ -32,7 +32,7 @@ class Channel(DefaultBase):
     title: Mapped[str]
     description: Mapped[str]
     deployment_id: Mapped[str] = mapped_column(unique=True)
-    llm_model: Mapped[str] = mapped_column(default=langchain_settings.default_model.value)
+    llm_model: Mapped[str] = mapped_column(default=langchain_settings.embedding_default_model.value)
     details: Mapped[dict[str, Any]] = mapped_column(type_=postgresql.JSONB)
 
     # ~~~~~ Relationships ~~~~~
@@ -150,15 +150,17 @@ class ChannelDatasetVersion(DefaultBase):
     __tablename__ = "channel_dataset_versions"
     __table_args__ = (
         UniqueConstraint(
-            'channel_dataset_id', 'version', name='uix_unique_version_for_channel_dataset'
+            "channel_dataset_id",
+            "version",
+            name="uix_unique_version_for_channel_dataset",
         ),
     )
 
     channel_dataset_id: Mapped[int] = mapped_column(ForeignKey("channel_datasets.id"))
     version: Mapped[int] = mapped_column(default=0)  # will be auto-incremented by trigger
     preprocessing_status: Mapped[PreprocessingStatusEnum]
     pointer_to: Mapped[int | None] = mapped_column(
-        ForeignKey("channel_dataset_versions.id", ondelete='SET NULL'), default=None
+        ForeignKey("channel_dataset_versions.id", ondelete="SET NULL"), default=None
     )
 
     creation_reason: Mapped[str]
@@ -181,14 +183,14 @@ class ChannelDatasetVersion(DefaultBase):
     channel_dataset: Mapped[ChannelDataset] = relationship(back_populates="versions")
     pointer = relationship(
         "ChannelDatasetVersion",
-        remote_side='ChannelDatasetVersion.id',
+        remote_side="ChannelDatasetVersion.id",
         back_populates="pointing_versions",
         cascade="all",
         passive_deletes=True,
     )
     pointing_versions = relationship(
         "ChannelDatasetVersion",
-        remote_side='ChannelDatasetVersion.pointer_to',
+        remote_side="ChannelDatasetVersion.pointer_to",
         back_populates="pointer",
         cascade="all, delete-orphan",
         passive_deletes=True,

statgpt/common/schemas/model_config.py

@@ -1,6 +1,11 @@
-from pydantic import Field
+from pydantic import Field, model_validator
 
-from statgpt.common.config import EmbeddingModelsEnum, LLMModelsEnum
+from statgpt.common.config import (
+    EmbeddingModelsEnum,
+    LLMModelsEnum,
+    ReasoningEffortEnum,
+    VerbosityEnum,
+)
 from statgpt.common.settings.langchain import langchain_settings
 
 from .base import BaseYamlModel
@@ -10,7 +15,8 @@ class BaseModelConfig(BaseYamlModel):
     """Base config for LLM and embeddings models configs."""
 
     api_version: str = Field(
-        default=langchain_settings.default_api_version, description="API version for the model"
+        default=langchain_settings.default_api_version,
+        description="API version for the model",
     )
 
 
@@ -30,16 +36,47 @@ class LLMModelConfig(BaseModelConfig):
         default=langchain_settings.default_model,
         description="The deployment of the model in DIAL",
     )
-    temperature: float = Field(
+    temperature: float | None = Field(
         default=langchain_settings.default_temperature,
         description=(
             "The temperature of the model. 0.0 means deterministic output, higher values mean more"
-            " randomness."
+            " randomness. Note: For reasoning models (except reasoning_effort=none) should be set to 1"
         ),
     )
     seed: int | None = Field(
         default=langchain_settings.default_seed,
         description=(
-            "The seed of the model. If set, the model will produce the same output for the same input."
+            "The seed of the model. If set, the model will produce the same output for the same input. "
+            "Note: Not supported by GPT-5 models."
         ),
     )
+    reasoning_effort: ReasoningEffortEnum | None = Field(
+        default=None,
+        description=(
+            "Reasoning effort level for GPT-5 models. "
+            "Supports: none, minimal, low, medium, high, xhigh. "
+            "All models before gpt-5.1 default to medium reasoning effort, and do not support none."
+        ),
+    )
+    verbosity: VerbosityEnum | None = Field(
+        default=None,
+        description="Output verbosity for GPT-5 models (low/medium/high).",
+    )
+
+    @model_validator(mode="after")
+    def _validate_model_family_params(self) -> "LLMModelConfig":
+        if self.deployment.is_gpt_5_family:
+            if self.seed is not None:
+                raise ValueError("seed is not supported for GPT-5 models")
+            if self.reasoning_effort is None:
+                raise ValueError("reasoning_effort is required for GPT-5 models")
+            if self.reasoning_effort is not ReasoningEffortEnum.NONE and self.temperature != 1:
+                raise ValueError(
+                    "temperature must be set to 1 when reasoning_effort is enabled for GPT-5 models"
+                )
+        else:
+            if self.reasoning_effort is not None:
+                raise ValueError("reasoning_effort is only supported for GPT-5 models")
+            if self.verbosity is not None:
+                raise ValueError("verbosity is only supported for GPT-5 models")
+        return self

statgpt/common/settings/langchain.py

@@ -1,6 +1,4 @@
-from typing import Optional
-
-from langchain import globals as lc_globals
+from langchain_core import globals as lc_globals
 from pydantic import Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
@@ -35,7 +33,7 @@ class LangChainSettings(BaseSettings):
         description="Default API version for Azure OpenAI",
     )
 
-    default_seed: Optional[int] = Field(
+    default_seed: int | None = Field(
         default=None,
         description="Default seed for reproducible outputs",
     )

statgpt/common/utils/models.py

@@ -15,9 +15,7 @@ def get_chat_model(
     model_config: LLMModelConfig,
     azure_endpoint: str = dial_settings.url,
     timeout: httpx.Timeout | None = None,
-    **kwargs,
 ) -> AzureChatOpenAI:
-    # default params
     if not isinstance(api_key, SecretStr):
         api_key = SecretStr(api_key)
     if not timeout:
@@ -26,13 +24,12 @@ def get_chat_model(
         azure_endpoint=azure_endpoint,
         api_version=model_config.api_version,
         azure_deployment=model_config.deployment.deployment_id,
-        temperature=model_config.temperature,
-        seed=model_config.seed,
         max_retries=10,
         api_key=api_key,  # since we use SecretStr, it won't be logged
         timeout=timeout,  # timeouts are crucial!
     )
-    params.update(kwargs)  # update default params
+
+    params.update(model_config.model_dump(mode="json", exclude_none=True, exclude={"deployment"}))
 
     if model_config.deployment.is_gpt_41_family:
         callback = BrokenResponseInterceptor(regex_pattern=r'\s{5,}')
@@ -49,7 +46,6 @@ def get_embeddings_model(
     api_key: str | SecretStr,
     model_config: EmbeddingsModelConfig,
     azure_endpoint: str = dial_settings.url,
-    **kwargs,
 ) -> AzureOpenAIEmbeddings:
     if not isinstance(api_key, SecretStr):
         api_key = SecretStr(api_key)
@@ -60,7 +56,6 @@ def get_embeddings_model(
         max_retries=10,
         api_key=api_key,  # since we use SecretStr, it won't be logged
     )
-    params.update(kwargs)  # update default params
     api_key_log = f'{api_key.get_secret_value()[:3]}*****{api_key.get_secret_value()[-2:]}'
     logger.info(
         f'creating langchain embeddings with the following params: {params}, Api key: {api_key_log}'