Feat: strip sampling params when modelDetails.shouldSkipTemperature is true (#74)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: cosminacho

CHANGELOG.md

@@ -2,6 +2,11 @@
 
 All notable changes to `uipath_llm_client` (core package) will be documented in this file.
 
+## [1.10.0] - 2026-04-23
+
+### Added
+- `uipath.llm_client.utils.sampling` module exposing `DISABLED_SAMPLING_PARAMS`, `disabled_params_from_model_details`, `is_disabled_value`, and `strip_disabled_kwargs`. The helpers use the langchain-openai-style `disabled_params` format (`{name: None | [values]}`) so they compose with the existing `langchain_openai._filter_disabled_params` path. `disabled_params_from_model_details` derives the disabled-param map from a discovery-endpoint `modelDetails` dict (today: `shouldSkipTemperature=True` disables the full sampling set — temperature, top_p, top_k, frequency/presence penalty, seed, logit_bias, logprobs, top_logprobs).
+
 ## [1.9.9] - 2026-04-23
 
 ### Changed

packages/uipath_langchain_client/CHANGELOG.md

@@ -2,6 +2,22 @@
 
 All notable changes to `uipath_langchain_client` will be documented in this file.
 
+## [1.10.0] - 2026-04-23
+
+### Added
+- `model_details` and `disabled_params` fields on `UiPathBaseLLMClient`, plus a single `@model_validator(mode="after") setup_model_info` that (1) forwards the factory-supplied `model_details` or fetches it from `client_settings.get_model_info`, and (2) sets `disabled_params` to the merge of what the caller passed and what `disabled_params_from_model_details` derives — user keys win on conflicts, so callers can override any derived entry by name.
+- `disabled_params` uses the langchain-openai shape (`{name: None | [values]}`), so subclasses inheriting from `ChatOpenAI` / `AzureChatOpenAI` also benefit from the native `_filter_disabled_params` path inside `bind_tools`.
+- Runtime stripping in the four `_generate`/`_agenerate`/`_stream`/`_astream` wrappers on `UiPathBaseChatModel` delegates to `uipath.llm_client.utils.sampling.strip_disabled_kwargs`, generic over `disabled_params`. A warning is logged via `self.logger` for each stripped key when a logger is configured. Fixes `anthropic.claude-opus-4-7` rejecting any sampling parameter passed via `.invoke()` / `.ainvoke()` / streams.
+
+### Removed
+- The unused `disabled_params` field declaration on `UiPathChat` (now inherited from `UiPathBaseLLMClient`).
+
+### Changed
+- Bumped `uipath-llm-client` floor to `>=1.10.0` to match the release that adds `uipath.llm_client.utils.sampling`.
+
+### Known follow-up
+- Init-time values set on the instance (`UiPathChat(model="anthropic.claude-opus-4-7", temperature=0.5)`) still flow into the outgoing request body via `_default_params` / the vendor SDK. The runtime invoke-time strip handles `.invoke(..., temperature=...)`; a follow-up will plug the init-time leak using the already-populated `disabled_params`.
+
 ## [1.9.9] - 2026-04-23
 
 ### Changed

packages/uipath_langchain_client/pyproject.toml

@@ -6,7 +6,7 @@ readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
     "langchain>=1.2.15,<2.0.0",
-    "uipath-llm-client>=1.9.9,<2.0.0",
+    "uipath-llm-client>=1.10.0,<2.0.0",
 ]
 
 [project.optional-dependencies]

packages/uipath_langchain_client/src/uipath_langchain_client/__version__.py

@@ -1,3 +1,3 @@
 __title__ = "UiPath LangChain Client"
 __description__ = "A Python client for interacting with UiPath's LLM services via LangChain."
-__version__ = "1.9.9"
+__version__ = "1.10.0"

packages/uipath_langchain_client/src/uipath_langchain_client/base_client.py

@@ -27,7 +27,7 @@
 from abc import ABC
 from collections.abc import AsyncGenerator, Generator, Mapping, Sequence
 from functools import cached_property
-from typing import Any, ClassVar, Literal
+from typing import Any, ClassVar, Literal, Self
 
 from httpx import URL, Response
 from langchain_core.callbacks import (
@@ -38,7 +38,7 @@
 from langchain_core.language_models.chat_models import BaseChatModel
 from langchain_core.messages import BaseMessage
 from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
-from pydantic import AliasChoices, BaseModel, ConfigDict, Field
+from pydantic import AliasChoices, BaseModel, ConfigDict, Field, model_validator
 
 from uipath.llm_client.httpx_client import (
     UiPathHttpxAsyncClient,
@@ -49,6 +49,10 @@
     get_captured_response_headers,
     set_captured_response_headers,
 )
+from uipath.llm_client.utils.sampling import (
+    disabled_params_from_model_details,
+    strip_disabled_kwargs,
+)
 from uipath_langchain_client.settings import (
     UiPathAPIConfig,
     UiPathBaseSettings,
@@ -108,6 +112,19 @@ class UiPathBaseLLMClient(BaseModel, ABC):
         description="Settings for the UiPath client (defaults based on UIPATH_LLM_SERVICE env var)",
     )
 
+    model_details: dict[str, Any] | None = Field(
+        default=None,
+        description="Per-model capability flags sourced from the discovery endpoint "
+        "(e.g. {'shouldSkipTemperature': True}). Passed through by the factory; "
+        "resolved from client_settings.get_model_info otherwise.",
+    )
+    disabled_params: dict[str, Any] | None = Field(
+        default=None,
+        description="langchain-openai-style map of parameters that must not be sent to "
+        "this model. Keys are param names; values are None (always disabled) or a list "
+        "of disallowed values. Derived from ``model_details`` when not provided.",
+    )
+
     default_headers: Mapping[str, str] | None = Field(
         default=None,
         description="Caller-supplied request headers. Merged on top of `class_default_headers`; "
@@ -140,6 +157,39 @@ class UiPathBaseLLMClient(BaseModel, ABC):
         description="Logger for request/response logging",
     )
 
+    @model_validator(mode="after")
+    def setup_model_info(self) -> Self:
+        """Resolve ``model_details`` from discovery and merge ``disabled_params``.
+
+        Runs after pydantic has validated the fields, so ``self.client_settings``
+        (with its ``default_factory``) and ``self.model_name`` are already live.
+
+        ``model_details`` is resolved once: caller-forwarded value wins, then a
+        lookup against ``client_settings.get_model_info`` (backed by the
+        class-cached discovery response), else an empty mapping on failure.
+
+        ``disabled_params`` is the merge of what the caller passed and what we
+        can derive from ``model_details`` (via
+        ``disabled_params_from_model_details``). User-provided keys win on
+        conflicts, so callers can override a derived entry by name.
+        """
+        if self.model_details is None:
+            try:
+                info = self.client_settings.get_model_info(
+                    self.model_name,
+                    byo_connection_id=self.byo_connection_id,
+                )
+                self.model_details = info.get("modelDetails") or {}
+            except Exception:
+                self.model_details = {}
+
+        derived = disabled_params_from_model_details(self.model_details) or {}
+        user_provided = self.disabled_params or {}
+        merged = {**derived, **user_provided}
+        self.disabled_params = merged or None
+
+        return self
+
     @cached_property
     def uipath_sync_client(self) -> UiPathHttpxClient:
         """Here we instantiate a synchronous HTTP client with the proper authentication pipeline, retry logic, logging etc."""
@@ -364,6 +414,12 @@ def _generate(
         run_manager: CallbackManagerForLLMRun | None = None,
         **kwargs: Any,
     ) -> ChatResult:
+        kwargs = strip_disabled_kwargs(
+            kwargs,
+            disabled_params=self.disabled_params,
+            model_name=self.model_name,
+            logger=self.logger,
+        )
         set_captured_response_headers({})
         try:
             result = self._uipath_generate(messages, stop=stop, run_manager=run_manager, **kwargs)
@@ -389,6 +445,12 @@ async def _agenerate(
         run_manager: AsyncCallbackManagerForLLMRun | None = None,
         **kwargs: Any,
     ) -> ChatResult:
+        kwargs = strip_disabled_kwargs(
+            kwargs,
+            disabled_params=self.disabled_params,
+            model_name=self.model_name,
+            logger=self.logger,
+        )
         set_captured_response_headers({})
         try:
             result = await self._uipath_agenerate(
@@ -416,6 +478,12 @@ def _stream(
         run_manager: CallbackManagerForLLMRun | None = None,
         **kwargs: Any,
     ) -> Generator[ChatGenerationChunk, None, None]:
+        kwargs = strip_disabled_kwargs(
+            kwargs,
+            disabled_params=self.disabled_params,
+            model_name=self.model_name,
+            logger=self.logger,
+        )
         set_captured_response_headers({})
         try:
             first = True
@@ -446,6 +514,12 @@ async def _astream(
         run_manager: AsyncCallbackManagerForLLMRun | None = None,
         **kwargs: Any,
     ) -> AsyncGenerator[ChatGenerationChunk, None]:
+        kwargs = strip_disabled_kwargs(
+            kwargs,
+            disabled_params=self.disabled_params,
+            model_name=self.model_name,
+            logger=self.logger,
+        )
         set_captured_response_headers({})
         try:
             first = True

packages/uipath_langchain_client/src/uipath_langchain_client/clients/litellm/embeddings.py

@@ -10,8 +10,6 @@
     >>> vectors = embeddings.embed_documents(["Hello world"])
 """
 
-from __future__ import annotations
-
 from pydantic import Field, model_validator
 from typing_extensions import Self
 

packages/uipath_langchain_client/src/uipath_langchain_client/clients/normalized/chat_models.py

@@ -64,9 +64,9 @@
 from langchain_core.utils.pydantic import is_basemodel_subclass
 from pydantic import AliasChoices, BaseModel, Field
 
+from uipath.llm_client.utils.model_family import is_anthropic_model_name
 from uipath_langchain_client.base_client import UiPathBaseChatModel
 from uipath_langchain_client.settings import ApiType, RoutingMode, UiPathAPIConfig
-from uipath_langchain_client.utils import is_anthropic_model_name
 
 _DictOrPydanticClass = Union[dict[str, Any], type[BaseModel], type]
 _DictOrPydantic = Union[dict[str, Any], BaseModel]
@@ -179,7 +179,6 @@ class UiPathChat(UiPathBaseChatModel):
     seed: int | None = None
 
     model_kwargs: dict[str, Any] = Field(default_factory=dict)
-    disabled_params: dict[str, Any] | None = None
 
     # OpenAI
     logit_bias: dict[str, int] | None = None

packages/uipath_langchain_client/src/uipath_langchain_client/factory.py

@@ -22,6 +22,7 @@
 
 from typing import Any
 
+from uipath.llm_client.utils.model_family import is_anthropic_model_name
 from uipath_langchain_client.base_client import (
     UiPathBaseChatModel,
     UiPathBaseEmbeddings,
@@ -36,7 +37,6 @@
     VendorType,
     get_default_client_settings,
 )
-from uipath_langchain_client.utils import is_anthropic_model_name
 
 
 def get_chat_model(
@@ -85,12 +85,14 @@ def get_chat_model(
         vendor_type=vendor_type,
     )
     model_family = model_info.get("modelFamily", None)
+    model_details = model_info.get("modelDetails") or {}
 
     if custom_class is not None:
         return custom_class(
             model=model_name,
             settings=client_settings,
             byo_connection_id=byo_connection_id,
+            model_details=model_details,
             **model_kwargs,
         )
 
@@ -103,6 +105,7 @@ def get_chat_model(
             model=model_name,
             settings=client_settings,
             byo_connection_id=byo_connection_id,
+            model_details=model_details,
             **model_kwargs,
         )
 
@@ -138,6 +141,7 @@ def get_chat_model(
                     settings=client_settings,
                     api_flavor=api_flavor,
                     byo_connection_id=byo_connection_id,
+                    model_details=model_details,
                     **model_kwargs,
                 )
             else:
@@ -150,6 +154,7 @@ def get_chat_model(
                     settings=client_settings,
                     api_flavor=api_flavor,
                     byo_connection_id=byo_connection_id,
+                    model_details=model_details,
                     **model_kwargs,
                 )
         case VendorType.VERTEXAI:
@@ -163,6 +168,7 @@ def get_chat_model(
                     settings=client_settings,
                     vendor_type=discovered_vendor_type,
                     byo_connection_id=byo_connection_id,
+                    model_details=model_details,
                     **model_kwargs,
                 )
 
@@ -174,6 +180,7 @@ def get_chat_model(
                 model=model_name,
                 settings=client_settings,
                 byo_connection_id=byo_connection_id,
+                model_details=model_details,
                 **model_kwargs,
             )
         case VendorType.AWSBEDROCK:
@@ -188,6 +195,7 @@ def get_chat_model(
                     model=model_name,
                     settings=client_settings,
                     byo_connection_id=byo_connection_id,
+                    model_details=model_details,
                     **model_kwargs,
                 )
 
@@ -200,6 +208,7 @@ def get_chat_model(
                     model=model_name,
                     settings=client_settings,
                     byo_connection_id=byo_connection_id,
+                    model_details=model_details,
                     **model_kwargs,
                 )
 
@@ -211,6 +220,7 @@ def get_chat_model(
                 model=model_name,
                 settings=client_settings,
                 byo_connection_id=byo_connection_id,
+                model_details=model_details,
                 **model_kwargs,
             )
 

packages/uipath_langchain_client/src/uipath_langchain_client/utils.py

@@ -13,10 +13,6 @@
     UiPathTooManyRequestsError,
     UiPathUnprocessableEntityError,
 )
-from uipath.llm_client.utils.model_family import (
-    ANTHROPIC_MODEL_NAME_KEYWORDS,
-    is_anthropic_model_name,
-)
 from uipath.llm_client.utils.retry import RetryConfig
 
 __all__ = [
@@ -34,6 +30,4 @@
     "UiPathServiceUnavailableError",
     "UiPathGatewayTimeoutError",
     "UiPathTooManyRequestsError",
-    "ANTHROPIC_MODEL_NAME_KEYWORDS",
-    "is_anthropic_model_name",
 ]

src/uipath/llm_client/__version__.py

@@ -1,3 +1,3 @@
 __title__ = "UiPath LLM Client"
 __description__ = "A Python client for interacting with UiPath's LLM services."
-__version__ = "1.9.9"
+__version__ = "1.10.0"