[Feat] - Display thinking tokens on OpenWebUI (Bedrock, Anthropic, Deepseek)

litellm/__init__.py

@@ -277,8 +277,6 @@
 custom_prometheus_metadata_labels: List[str] = []
 #### REQUEST PRIORITIZATION ####
 priority_reservation: Optional[Dict[str, float]] = None
-
-
 force_ipv4: bool = (
     False  # when True, litellm will force ipv4 for all LLM requests. Some users have seen httpx ConnectionError when using ipv6.
 )

litellm/litellm_core_utils/get_litellm_params.py

@@ -57,6 +57,7 @@ def get_litellm_params(
     prompt_variables: Optional[dict] = None,
     async_call: Optional[bool] = None,
     ssl_verify: Optional[bool] = None,
+    merge_reasoning_content_in_choices: Optional[bool] = None,
     **kwargs,
 ) -> dict:
     litellm_params = {
@@ -97,5 +98,6 @@ def get_litellm_params(
         "prompt_variables": prompt_variables,
         "async_call": async_call,
         "ssl_verify": ssl_verify,
+        "merge_reasoning_content_in_choices": merge_reasoning_content_in_choices,
     }
     return litellm_params

litellm/litellm_core_utils/streaming_handler.py

@@ -15,6 +15,7 @@
 from litellm.litellm_core_utils.redact_messages import LiteLLMLoggingObject
 from litellm.litellm_core_utils.thread_pool_executor import executor
 from litellm.types.llms.openai import ChatCompletionChunk
+from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import Delta
 from litellm.types.utils import GenericStreamingChunk as GChunk
 from litellm.types.utils import (
@@ -70,6 +71,17 @@ def __init__(
         self.completion_stream = completion_stream
         self.sent_first_chunk = False
         self.sent_last_chunk = False
+
+        litellm_params: GenericLiteLLMParams = GenericLiteLLMParams(
+            **self.logging_obj.model_call_details.get("litellm_params", {})
+        )
+        self.merge_reasoning_content_in_choices: bool = (
+            litellm_params.merge_reasoning_content_in_choices or False
+        )
+        self.sent_first_thinking_block = False
+        self.sent_last_thinking_block = False
+        self.thinking_content = ""
+
         self.system_fingerprint: Optional[str] = None
         self.received_finish_reason: Optional[str] = None
         self.intermittent_finish_reason: Optional[str] = (
@@ -87,12 +99,7 @@ def __init__(
         self.holding_chunk = ""
         self.complete_response = ""
         self.response_uptil_now = ""
-        _model_info = (
-            self.logging_obj.model_call_details.get("litellm_params", {}).get(
-                "model_info", {}
-            )
-            or {}
-        )
+        _model_info: Dict = litellm_params.model_info or {}
 
         _api_base = get_api_base(
             model=model or "",
@@ -873,6 +880,10 @@ def return_processed_chunk_logic(  # noqa
                     _index: Optional[int] = completion_obj.get("index")
                     if _index is not None:
                         model_response.choices[0].index = _index
+
+                self._optional_combine_thinking_block_in_choices(
+                    model_response=model_response
+                )
                 print_verbose(f"returning model_response: {model_response}")
                 return model_response
             else:
@@ -929,6 +940,48 @@ def return_processed_chunk_logic(  # noqa
                 self.chunks.append(model_response)
             return
 
+    def _optional_combine_thinking_block_in_choices(
+        self, model_response: ModelResponseStream
+    ) -> None:
+        """
+        UI's Like OpenWebUI expect to get 1 chunk with <think>...</think> tags in the chunk content
+
+        In place updates the model_response object with reasoning_content in content with <think>...</think> tags
+
+        Enabled when `merge_reasoning_content_in_choices=True` passed in request params
+
+
+        """
+        if self.merge_reasoning_content_in_choices is True:
+            reasoning_content = getattr(
+                model_response.choices[0].delta, "reasoning_content", None
+            )
+            if reasoning_content:
+                if self.sent_first_thinking_block is False:
+                    model_response.choices[0].delta.content += (
+                        "<think>" + reasoning_content
+                    )
+                    self.sent_first_thinking_block = True
+                elif (
+                    self.sent_first_thinking_block is True
+                    and hasattr(model_response.choices[0].delta, "reasoning_content")
+                    and model_response.choices[0].delta.reasoning_content
+                ):
+                    model_response.choices[0].delta.content = reasoning_content
+            elif (
+                self.sent_first_thinking_block is True
+                and not self.sent_last_thinking_block
+                and model_response.choices[0].delta.content
+            ):
+                model_response.choices[0].delta.content = (
+                    "</think>" + model_response.choices[0].delta.content
+                )
+                self.sent_last_thinking_block = True
+
+            if hasattr(model_response.choices[0].delta, "reasoning_content"):
+                del model_response.choices[0].delta.reasoning_content
+        return
+
     def chunk_creator(self, chunk: Any):  # type: ignore  # noqa: PLR0915
         model_response = self.model_response_creator()
         response_obj: Dict[str, Any] = {}

litellm/main.py

@@ -1159,6 +1159,9 @@ def completion(  # type: ignore # noqa: PLR0915
             prompt_id=prompt_id,
             prompt_variables=prompt_variables,
             ssl_verify=ssl_verify,
+            merge_reasoning_content_in_choices=kwargs.get(
+                "merge_reasoning_content_in_choices", None
+            ),
         )
         logging.update_environment_variables(
             model=model,

litellm/proxy/proxy_config.yaml

@@ -1,27 +1,13 @@
 model_list:
-  - model_name: fake-openai-endpoint
+  - model_name: bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0
     litellm_params:
-      model: openai/my-fake-model
-      api_key: my-fake-key
-      api_base: https://exampleopenaiendpoint-production.up.railway.app/
-  - model_name: claude-special-alias
-    litellm_params:
-      model: anthropic/claude-3-haiku-20240307
-      api_key: os.environ/ANTHROPIC_API_KEY
-  - model_name: claude-3-5-sonnet-20241022
-    litellm_params:
-      model: anthropic/claude-3-5-sonnet-20241022
-      api_key: os.environ/ANTHROPIC_API_KEY
-  - model_name: claude-3-7-sonnet-20250219
-    litellm_params:
-      model: anthropic/claude-3-7-sonnet-20250219
-      api_key: os.environ/ANTHROPIC_API_KEY
-  - model_name: anthropic/*
-    litellm_params:
-      model: anthropic/*
-      api_key: os.environ/ANTHROPIC_API_KEY
+
+      model: bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0
+      thinking: {"type": "enabled", "budget_tokens": 1024}
+      max_tokens: 1080
+      merge_reasoning_content_in_choices: true
+
 
 general_settings:
   store_model_in_db: true
   store_prompts_in_spend_logs: true
-

litellm/types/router.py

@@ -192,6 +192,8 @@ class GenericLiteLLMParams(BaseModel):
     budget_duration: Optional[str] = None
     use_in_pass_through: Optional[bool] = False
     model_config = ConfigDict(extra="allow", arbitrary_types_allowed=True)
+    merge_reasoning_content_in_choices: Optional[bool] = False
+    model_info: Optional[Dict] = None
 
     def __init__(
         self,
@@ -231,6 +233,9 @@ def __init__(
         budget_duration: Optional[str] = None,
         # Pass through params
         use_in_pass_through: Optional[bool] = False,
+        # This will merge the reasoning content in the choices
+        merge_reasoning_content_in_choices: Optional[bool] = False,
+        model_info: Optional[Dict] = None,
         **params,
     ):
         args = locals()

litellm/types/utils.py

@@ -21,6 +21,8 @@
 from pydantic import BaseModel, ConfigDict, Field, PrivateAttr
 from typing_extensions import Callable, Dict, Required, TypedDict, override
 
+import litellm
+
 from ..litellm_core_utils.core_helpers import map_finish_reason
 from .guardrails import GuardrailEventHooks
 from .llms.openai import (
@@ -1803,6 +1805,7 @@ class StandardCallbackDynamicParams(TypedDict, total=False):
     "max_budget",
     "budget_duration",
     "use_in_pass_through",
+    "merge_reasoning_content_in_choices",
 ] + list(StandardCallbackDynamicParams.__annotations__.keys())