From 766b300fbc69f2502b85cbd8587fc9e94788dae3 Mon Sep 17 00:00:00 2001
From: oryx1729 <oryx1729@protonmail.com>
Date: Sat, 20 Sep 2025 16:35:51 -0700
Subject: [PATCH] fix: handle reasoning_effort parameter gracefully across
 models

- Set litellm.drop_params=True to auto-drop unsupported parameters
- Changed reasoning_effort from extra_body to direct parameter
- Added reasoning_effort to both async and sync methods
- Removed redundant retry logic for unsupported parameters
- Ensures compatibility with models that don't support reasoning_effort

This fixes errors when using models that don't support the reasoning_effort
parameter while maintaining the optimization for models that do support it.
---
 .../litellm_instructor/llm/openai/adapter.py  | 31 +++----------------
 1 file changed, 4 insertions(+), 27 deletions(-)

diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py
index 7e8d75d10..12dad5f55 100644
--- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py
+++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py
@@ -29,6 +29,8 @@ observe = get_observe()
 
 logger = get_logger()
 
+# litellm to drop unsupported params, e.g., reasoning_effort when not supported by the model.
+litellm.drop_params = True
 
 class OpenAIAdapter(LLMInterface):
     """
@@ -132,39 +134,13 @@ class OpenAIAdapter(LLMInterface):
                 api_version=self.api_version,
                 response_model=response_model,
                 max_retries=self.MAX_RETRIES,
-                extra_body={"reasoning_effort": "minimal"},
+                reasoning_effort="minimal",
             )
         except (
             ContentFilterFinishReasonError,
             ContentPolicyViolationError,
             InstructorRetryException,
         ) as error:
-            if (
-                isinstance(error, InstructorRetryException)
-                and "content management policy" not in str(error).lower()
-            ):
-                logger.debug(
-                    "LLM Model does not support reasoning_effort parameter, trying call without the parameter."
-                )
-                return await self.aclient.chat.completions.create(
-                    model=self.model,
-                    messages=[
-                        {
-                            "role": "user",
-                            "content": f"""{text_input}""",
-                        },
-                        {
-                            "role": "system",
-                            "content": system_prompt,
-                        },
-                    ],
-                    api_key=self.api_key,
-                    api_base=self.endpoint,
-                    api_version=self.api_version,
-                    response_model=response_model,
-                    max_retries=self.MAX_RETRIES,
-                )
-
             if not (self.fallback_model and self.fallback_api_key):
                 raise ContentPolicyFilterError(
                     f"The provided input contains content that is not aligned with our content policy: {text_input}"
@@ -246,6 +222,7 @@ class OpenAIAdapter(LLMInterface):
             api_base=self.endpoint,
             api_version=self.api_version,
             response_model=response_model,
+            reasoning_effort="minimal",
             max_retries=self.MAX_RETRIES,
         )