agno 2.3.11__py3-none-any.whl → 2.3.12__py3-none-any.whl

agno/models/base.py

@@ -8,6 +8,7 @@ from pathlib import Path
 from time import sleep, time
 from types import AsyncGeneratorType, GeneratorType
 from typing import (
+    TYPE_CHECKING,
     Any,
     AsyncIterator,
     Dict,
@@ -15,11 +16,15 @@ from typing import (
     List,
     Literal,
     Optional,
+    Sequence,
     Tuple,
     Type,
     Union,
     get_args,
 )
+
+if TYPE_CHECKING:
+    from agno.compression.manager import CompressionManager
 from uuid import uuid4
 
 from pydantic import BaseModel
@@ -156,6 +161,8 @@ class Model(ABC):
     # Enable retrying a model invocation once with a guidance message.
     # This is useful for known errors avoidable with extra instructions.
     retry_with_guidance: bool = True
+    # Set the number of times to retry the model invocation with guidance.
+    retry_with_guidance_limit: int = 1
 
     def __post_init__(self):
         if self.provider is None and self.name is not None:
@@ -178,6 +185,7 @@ class Model(ABC):
 
         for attempt in range(self.retries + 1):
             try:
+                retries_with_guidance_count = kwargs.pop("retries_with_guidance_count", 0)
                 return self.invoke(**kwargs)
             except ModelProviderError as e:
                 last_exception = e
@@ -190,8 +198,20 @@ class Model(ABC):
                 else:
                     log_error(f"Model provider error after {self.retries + 1} attempts: {e}")
             except RetryableModelProviderError as e:
+                current_count = retries_with_guidance_count
+                if current_count >= self.retry_with_guidance_limit:
+                    raise ModelProviderError(
+                        message=f"Max retries with guidance reached. Error: {e.original_error}",
+                        model_name=self.name,
+                        model_id=self.id,
+                    )
+                kwargs.pop("retry_with_guidance", None)
+                kwargs["retries_with_guidance_count"] = current_count + 1
+
+                # Append the guidance message to help the model avoid the error in the next invoke.
                 kwargs["messages"].append(Message(role="user", content=e.retry_guidance_message, temporary=True))
-                return self._invoke_with_retry(**kwargs, retrying_with_guidance=True)
+
+                return self._invoke_with_retry(**kwargs, retry_with_guidance=True)
 
         # If we've exhausted all retries, raise the last exception
         raise last_exception  # type: ignore
@@ -207,6 +227,7 @@ class Model(ABC):
 
         for attempt in range(self.retries + 1):
             try:
+                retries_with_guidance_count = kwargs.pop("retries_with_guidance_count", 0)
                 return await self.ainvoke(**kwargs)
             except ModelProviderError as e:
                 last_exception = e
@@ -219,8 +240,21 @@ class Model(ABC):
                 else:
                     log_error(f"Model provider error after {self.retries + 1} attempts: {e}")
             except RetryableModelProviderError as e:
+                current_count = retries_with_guidance_count
+                if current_count >= self.retry_with_guidance_limit:
+                    raise ModelProviderError(
+                        message=f"Max retries with guidance reached. Error: {e.original_error}",
+                        model_name=self.name,
+                        model_id=self.id,
+                    )
+
+                kwargs.pop("retry_with_guidance", None)
+                kwargs["retries_with_guidance_count"] = current_count + 1
+
+                # Append the guidance message to help the model avoid the error in the next invoke.
                 kwargs["messages"].append(Message(role="user", content=e.retry_guidance_message, temporary=True))
-                return await self._ainvoke_with_retry(**kwargs, retrying_with_guidance=True)
+
+                return await self._ainvoke_with_retry(**kwargs, retry_with_guidance=True)
 
         # If we've exhausted all retries, raise the last exception
         raise last_exception  # type: ignore
@@ -236,6 +270,7 @@ class Model(ABC):
 
         for attempt in range(self.retries + 1):
             try:
+                retries_with_guidance_count = kwargs.pop("retries_with_guidance_count", 0)
                 yield from self.invoke_stream(**kwargs)
                 return  # Success, exit the retry loop
             except ModelProviderError as e:
@@ -250,8 +285,21 @@ class Model(ABC):
                 else:
                     log_error(f"Model provider error after {self.retries + 1} attempts: {e}")
             except RetryableModelProviderError as e:
+                current_count = retries_with_guidance_count
+                if current_count >= self.retry_with_guidance_limit:
+                    raise ModelProviderError(
+                        message=f"Max retries with guidance reached. Error: {e.original_error}",
+                        model_name=self.name,
+                        model_id=self.id,
+                    )
+
+                kwargs.pop("retry_with_guidance", None)
+                kwargs["retries_with_guidance_count"] = current_count + 1
+
+                # Append the guidance message to help the model avoid the error in the next invoke.
                 kwargs["messages"].append(Message(role="user", content=e.retry_guidance_message, temporary=True))
-                yield from self._invoke_stream_with_retry(**kwargs, retrying_with_guidance=True)
+
+                yield from self._invoke_stream_with_retry(**kwargs, retry_with_guidance=True)
                 return  # Success, exit after regeneration
 
         # If we've exhausted all retries, raise the last exception
@@ -268,6 +316,7 @@ class Model(ABC):
 
         for attempt in range(self.retries + 1):
             try:
+                retries_with_guidance_count = kwargs.pop("retries_with_guidance_count", 0)
                 async for response in self.ainvoke_stream(**kwargs):
                     yield response
                 return  # Success, exit the retry loop
@@ -283,8 +332,21 @@ class Model(ABC):
                 else:
                     log_error(f"Model provider error after {self.retries + 1} attempts: {e}")
             except RetryableModelProviderError as e:
+                current_count = retries_with_guidance_count
+                if current_count >= self.retry_with_guidance_limit:
+                    raise ModelProviderError(
+                        message=f"Max retries with guidance reached. Error: {e.original_error}",
+                        model_name=self.name,
+                        model_id=self.id,
+                    )
+
+                kwargs.pop("retry_with_guidance", None)
+                kwargs["retries_with_guidance_count"] = current_count + 1
+
+                # Append the guidance message to help the model avoid the error in the next invoke.
                 kwargs["messages"].append(Message(role="user", content=e.retry_guidance_message, temporary=True))
-                async for response in self._ainvoke_stream_with_retry(**kwargs, retrying_with_guidance=True):
+
+                async for response in self._ainvoke_stream_with_retry(**kwargs, retry_with_guidance=True):
                     yield response
                 return  # Success, exit after regeneration
 
@@ -296,8 +358,8 @@ class Model(ABC):
         _dict = {field: getattr(self, field) for field in fields if getattr(self, field) is not None}
         return _dict
 
-    def _remove_temporarys(self, messages: List[Message]) -> None:
-        """Remove temporal messages from the given list.
+    def _remove_temporary_messages(self, messages: List[Message]) -> None:
+        """Remove temporary messages from the given list.
 
         Args:
             messages: The list of messages to filter (modified in place).
@@ -453,6 +515,29 @@ class Model(ABC):
                 _tool_dicts.append(tool)
         return _tool_dicts
 
+    def count_tokens(
+        self,
+        messages: List[Message],
+        tools: Optional[Sequence[Union[Function, Dict[str, Any]]]] = None,
+        output_schema: Optional[Union[Dict, Type[BaseModel]]] = None,
+    ) -> int:
+        from agno.utils.tokens import count_tokens
+
+        return count_tokens(
+            messages,
+            tools=list(tools) if tools else None,
+            model_id=self.id,
+            output_schema=output_schema,
+        )
+
+    async def acount_tokens(
+        self,
+        messages: List[Message],
+        tools: Optional[Sequence[Union[Function, Dict[str, Any]]]] = None,
+        output_schema: Optional[Union[Dict, Type[BaseModel]]] = None,
+    ) -> int:
+        return self.count_tokens(messages, tools, output_schema=output_schema)
+
     def response(
         self,
         messages: List[Message],
@@ -462,7 +547,7 @@ class Model(ABC):
         tool_call_limit: Optional[int] = None,
         run_response: Optional[Union[RunOutput, TeamRunOutput]] = None,
         send_media_to_model: bool = True,
-        compression_manager: Optional[Any] = None,
+        compression_manager: Optional["CompressionManager"] = None,
     ) -> ModelResponse:
         """
         Generate a response from the model.
@@ -500,8 +585,15 @@ class Model(ABC):
             _functions = {tool.name: tool for tool in tools if isinstance(tool, Function)} if tools is not None else {}
 
             _compress_tool_results = compression_manager is not None and compression_manager.compress_tool_results
+            _compression_manager = compression_manager if _compress_tool_results else None
 
             while True:
+                # Compress tool results if compression is enabled and threshold is met
+                if _compression_manager is not None and _compression_manager.should_compress(
+                    messages, tools, model=self, response_format=response_format
+                ):
+                    _compression_manager.compress(messages)
+
                 # Get response from model
                 assistant_message = Message(role=self.assistant_message_role)
                 self._process_model_response(
@@ -600,11 +692,6 @@ class Model(ABC):
                     # Add a function call for each successful execution
                     function_call_count += len(function_call_results)
 
-                    all_messages = messages + function_call_results
-                    # Compress tool results
-                    if compression_manager and compression_manager.should_compress(all_messages):
-                        compression_manager.compress(all_messages)
-
                     # Format and add results to messages
                     self.format_function_call_results(
                         messages=messages,
@@ -674,11 +761,12 @@ class Model(ABC):
         tool_call_limit: Optional[int] = None,
         run_response: Optional[Union[RunOutput, TeamRunOutput]] = None,
         send_media_to_model: bool = True,
-        compression_manager: Optional[Any] = None,
+        compression_manager: Optional["CompressionManager"] = None,
     ) -> ModelResponse:
         """
         Generate an asynchronous response from the model.
         """
+
         try:
             # Check cache if enabled
             if self.cache_response:
@@ -700,10 +788,17 @@ class Model(ABC):
             _functions = {tool.name: tool for tool in tools if isinstance(tool, Function)} if tools is not None else {}
 
             _compress_tool_results = compression_manager is not None and compression_manager.compress_tool_results
+            _compression_manager = compression_manager if _compress_tool_results else None
 
             function_call_count = 0
 
             while True:
+                # Compress existing tool results BEFORE making API call to avoid context overflow
+                if _compression_manager is not None and await _compression_manager.ashould_compress(
+                    messages, tools, model=self, response_format=response_format
+                ):
+                    await _compression_manager.acompress(messages)
+
                 # Get response from model
                 assistant_message = Message(role=self.assistant_message_role)
                 await self._aprocess_model_response(
@@ -801,11 +896,6 @@ class Model(ABC):
                     # Add a function call for each successful execution
                     function_call_count += len(function_call_results)
 
-                    all_messages = messages + function_call_results
-                    # Compress tool results
-                    if compression_manager and compression_manager.should_compress(all_messages):
-                        await compression_manager.acompress(all_messages)
-
                     # Format and add results to messages
                     self.format_function_call_results(
                         messages=messages,
@@ -1093,7 +1183,7 @@ class Model(ABC):
         stream_model_response: bool = True,
         run_response: Optional[Union[RunOutput, TeamRunOutput]] = None,
         send_media_to_model: bool = True,
-        compression_manager: Optional[Any] = None,
+        compression_manager: Optional["CompressionManager"] = None,
     ) -> Iterator[Union[ModelResponse, RunOutputEvent, TeamRunOutputEvent]]:
         """
         Generate a streaming response from the model.
@@ -1127,10 +1217,17 @@ class Model(ABC):
             _functions = {tool.name: tool for tool in tools if isinstance(tool, Function)} if tools is not None else {}
 
             _compress_tool_results = compression_manager is not None and compression_manager.compress_tool_results
+            _compression_manager = compression_manager if _compress_tool_results else None
 
             function_call_count = 0
 
             while True:
+                # Compress existing tool results BEFORE invoke
+                if _compression_manager is not None and _compression_manager.should_compress(
+                    messages, tools, model=self, response_format=response_format
+                ):
+                    _compression_manager.compress(messages)
+
                 assistant_message = Message(role=self.assistant_message_role)
                 # Create assistant message and stream data
                 stream_data = MessageData()
@@ -1192,11 +1289,6 @@ class Model(ABC):
                     # Add a function call for each successful execution
                     function_call_count += len(function_call_results)
 
-                    all_messages = messages + function_call_results
-                    # Compress tool results
-                    if compression_manager and compression_manager.should_compress(all_messages):
-                        compression_manager.compress(all_messages)
-
                     # Format and add results to messages
                     if stream_data and stream_data.extra is not None:
                         self.format_function_call_results(
@@ -1311,7 +1403,7 @@ class Model(ABC):
         stream_model_response: bool = True,
         run_response: Optional[Union[RunOutput, TeamRunOutput]] = None,
         send_media_to_model: bool = True,
-        compression_manager: Optional[Any] = None,
+        compression_manager: Optional["CompressionManager"] = None,
     ) -> AsyncIterator[Union[ModelResponse, RunOutputEvent, TeamRunOutputEvent]]:
         """
         Generate an asynchronous streaming response from the model.
@@ -1345,10 +1437,17 @@ class Model(ABC):
             _functions = {tool.name: tool for tool in tools if isinstance(tool, Function)} if tools is not None else {}
 
             _compress_tool_results = compression_manager is not None and compression_manager.compress_tool_results
+            _compression_manager = compression_manager if _compress_tool_results else None
 
             function_call_count = 0
 
             while True:
+                # Compress existing tool results BEFORE making API call to avoid context overflow
+                if _compression_manager is not None and await _compression_manager.ashould_compress(
+                    messages, tools, model=self, response_format=response_format
+                ):
+                    await _compression_manager.acompress(messages)
+
                 # Create assistant message and stream data
                 assistant_message = Message(role=self.assistant_message_role)
                 stream_data = MessageData()
@@ -1410,11 +1509,6 @@ class Model(ABC):
                     # Add a function call for each successful execution
                     function_call_count += len(function_call_results)
 
-                    all_messages = messages + function_call_results
-                    # Compress tool results
-                    if compression_manager and compression_manager.should_compress(all_messages):
-                        await compression_manager.acompress(all_messages)
-
                     # Format and add results to messages
                     if stream_data and stream_data.extra is not None:
                         self.format_function_call_results(

agno/models/google/gemini.py

@@ -19,8 +19,10 @@ from agno.models.message import Citations, Message, UrlCitation
 from agno.models.metrics import Metrics
 from agno.models.response import ModelResponse
 from agno.run.agent import RunOutput
+from agno.tools.function import Function
 from agno.utils.gemini import format_function_definitions, format_image_for_message, prepare_response_schema
 from agno.utils.log import log_debug, log_error, log_info, log_warning
+from agno.utils.tokens import count_schema_tokens, count_text_tokens, count_tool_tokens
 
 try:
     from google import genai
@@ -310,6 +312,113 @@ class Gemini(Model):
             log_debug(f"Calling {self.provider} with request parameters: {request_params}", log_level=2)
         return request_params
 
+    def count_tokens(
+        self,
+        messages: List[Message],
+        tools: Optional[List[Union[Function, Dict[str, Any]]]] = None,
+        output_schema: Optional[Union[Dict, Type[BaseModel]]] = None,
+    ) -> int:
+        contents, system_instruction = self._format_messages(messages, compress_tool_results=True)
+        schema_tokens = count_schema_tokens(output_schema, self.id)
+
+        if self.vertexai:
+            # VertexAI supports full token counting with system_instruction and tools
+            config: Dict[str, Any] = {}
+            if system_instruction:
+                config["system_instruction"] = system_instruction
+            if tools:
+                formatted_tools = self._format_tools(tools)
+                gemini_tools = format_function_definitions(formatted_tools)
+                if gemini_tools:
+                    config["tools"] = [gemini_tools]
+
+            response = self.get_client().models.count_tokens(
+                model=self.id,
+                contents=contents,
+                config=config if config else None,  # type: ignore
+            )
+            return (response.total_tokens or 0) + schema_tokens
+        else:
+            # Google AI Studio: Use API for content tokens + local estimation for system/tools
+            # The API doesn't support system_instruction or tools in config, so we use a hybrid approach:
+            # 1. Get accurate token count for contents (text + multimodal) from API
+            # 2. Add estimated tokens for system_instruction and tools locally
+            try:
+                response = self.get_client().models.count_tokens(
+                    model=self.id,
+                    contents=contents,
+                )
+                total = response.total_tokens or 0
+            except Exception as e:
+                log_warning(f"Gemini count_tokens API failed: {e}. Falling back to tiktoken-based estimation.")
+                return super().count_tokens(messages, tools, output_schema)
+
+            # Add estimated tokens for system instruction (not supported by Google AI Studio API)
+            if system_instruction:
+                system_text = system_instruction if isinstance(system_instruction, str) else str(system_instruction)
+                total += count_text_tokens(system_text, self.id)
+
+            # Add estimated tokens for tools (not supported by Google AI Studio API)
+            if tools:
+                total += count_tool_tokens(tools, self.id)
+
+            # Add estimated tokens for response_format/output_schema
+            total += schema_tokens
+
+            return total
+
+    async def acount_tokens(
+        self,
+        messages: List[Message],
+        tools: Optional[List[Union[Function, Dict[str, Any]]]] = None,
+        output_schema: Optional[Union[Dict, Type[BaseModel]]] = None,
+    ) -> int:
+        contents, system_instruction = self._format_messages(messages, compress_tool_results=True)
+        schema_tokens = count_schema_tokens(output_schema, self.id)
+
+        # VertexAI supports full token counting with system_instruction and tools
+        if self.vertexai:
+            config: Dict[str, Any] = {}
+            if system_instruction:
+                config["system_instruction"] = system_instruction
+            if tools:
+                formatted_tools = self._format_tools(tools)
+                gemini_tools = format_function_definitions(formatted_tools)
+                if gemini_tools:
+                    config["tools"] = [gemini_tools]
+
+            response = await self.get_client().aio.models.count_tokens(
+                model=self.id,
+                contents=contents,
+                config=config if config else None,  # type: ignore
+            )
+            return (response.total_tokens or 0) + schema_tokens
+        else:
+            # Hybrid approach - Google AI Studio does not support system_instruction or tools in config
+            try:
+                response = await self.get_client().aio.models.count_tokens(
+                    model=self.id,
+                    contents=contents,
+                )
+                total = response.total_tokens or 0
+            except Exception as e:
+                log_warning(f"Gemini count_tokens API failed: {e}. Falling back to tiktoken-based estimation.")
+                return await super().acount_tokens(messages, tools, output_schema)
+
+            # Add estimated tokens for system instruction
+            if system_instruction:
+                system_text = system_instruction if isinstance(system_instruction, str) else str(system_instruction)
+                total += count_text_tokens(system_text, self.id)
+
+            # Add estimated tokens for tools
+            if tools:
+                total += count_tool_tokens(tools, self.id)
+
+            # Add estimated tokens for response_format/output_schema
+            total += schema_tokens
+
+            return total
+
     def invoke(
         self,
         messages: List[Message],
@@ -319,7 +428,7 @@ class Gemini(Model):
         tool_choice: Optional[Union[str, Dict[str, Any]]] = None,
         run_response: Optional[RunOutput] = None,
         compress_tool_results: bool = False,
-        retrying_with_guidance: bool = False,
+        retry_with_guidance: bool = False,
     ) -> ModelResponse:
         """
         Invokes the model with a list of messages and returns the response.
@@ -341,12 +450,12 @@ class Gemini(Model):
             assistant_message.metrics.stop_timer()
 
             model_response = self._parse_provider_response(
-                provider_response, response_format=response_format, retrying_with_guidance=retrying_with_guidance
+                provider_response, response_format=response_format, retry_with_guidance=retry_with_guidance
             )
 
             # If we were retrying the invoke with guidance, remove the guidance message
-            if retrying_with_guidance is True:
-                self._remove_temporarys(messages)
+            if retry_with_guidance is True:
+                self._remove_temporary_messages(messages)
 
             return model_response
 
@@ -374,7 +483,7 @@ class Gemini(Model):
         tool_choice: Optional[Union[str, Dict[str, Any]]] = None,
         run_response: Optional[RunOutput] = None,
         compress_tool_results: bool = False,
-        retrying_with_guidance: bool = False,
+        retry_with_guidance: bool = False,
     ) -> Iterator[ModelResponse]:
         """
         Invokes the model with a list of messages and returns the response as a stream.
@@ -394,11 +503,11 @@ class Gemini(Model):
                 contents=formatted_messages,
                 **request_kwargs,
             ):
-                yield self._parse_provider_response_delta(response, retrying_with_guidance=retrying_with_guidance)
+                yield self._parse_provider_response_delta(response, retry_with_guidance=retry_with_guidance)
 
             # If we were retrying the invoke with guidance, remove the guidance message
-            if retrying_with_guidance is True:
-                self._remove_temporarys(messages)
+            if retry_with_guidance is True:
+                self._remove_temporary_messages(messages)
 
             assistant_message.metrics.stop_timer()
 
@@ -425,7 +534,7 @@ class Gemini(Model):
         tool_choice: Optional[Union[str, Dict[str, Any]]] = None,
         run_response: Optional[RunOutput] = None,
         compress_tool_results: bool = False,
-        retrying_with_guidance: bool = False,
+        retry_with_guidance: bool = False,
     ) -> ModelResponse:
         """
         Invokes the model with a list of messages and returns the response.
@@ -449,12 +558,12 @@ class Gemini(Model):
             assistant_message.metrics.stop_timer()
 
             model_response = self._parse_provider_response(
-                provider_response, response_format=response_format, retrying_with_guidance=retrying_with_guidance
+                provider_response, response_format=response_format, retry_with_guidance=retry_with_guidance
             )
 
             # If we were retrying the invoke with guidance, remove the guidance message
-            if retrying_with_guidance is True:
-                self._remove_temporarys(messages)
+            if retry_with_guidance is True:
+                self._remove_temporary_messages(messages)
 
             return model_response
 
@@ -481,7 +590,7 @@ class Gemini(Model):
         tool_choice: Optional[Union[str, Dict[str, Any]]] = None,
         run_response: Optional[RunOutput] = None,
         compress_tool_results: bool = False,
-        retrying_with_guidance: bool = False,
+        retry_with_guidance: bool = False,
     ) -> AsyncIterator[ModelResponse]:
         """
         Invokes the model with a list of messages and returns the response as a stream.
@@ -504,11 +613,11 @@ class Gemini(Model):
                 **request_kwargs,
             )
             async for chunk in async_stream:
-                yield self._parse_provider_response_delta(chunk, retrying_with_guidance=retrying_with_guidance)
+                yield self._parse_provider_response_delta(chunk, retry_with_guidance=retry_with_guidance)
 
             # If we were retrying the invoke with guidance, remove the guidance message
-            if retrying_with_guidance is True:
-                self._remove_temporarys(messages)
+            if retry_with_guidance is True:
+                self._remove_temporary_messages(messages)
 
             assistant_message.metrics.stop_timer()
 
@@ -874,6 +983,8 @@ class Gemini(Model):
         """
         combined_original_content: List = []
         combined_function_result: List = []
+        tool_names: List[str] = []
+
         message_metrics = Metrics()
 
         if len(function_call_results) > 0:
@@ -883,13 +994,18 @@ class Gemini(Model):
                 combined_function_result.append(
                     {"tool_call_id": result.tool_call_id, "tool_name": result.tool_name, "content": compressed_content}
                 )
+                if result.tool_name:
+                    tool_names.append(result.tool_name)
                 message_metrics += result.metrics
 
+        tool_name = ", ".join(tool_names) if tool_names else None
+
         if combined_original_content:
             messages.append(
                 Message(
                     role="tool",
                     content=combined_original_content,
+                    tool_name=tool_name,
                     tool_calls=combined_function_result,
                     metrics=message_metrics,
                 )
@@ -915,11 +1031,11 @@ class Gemini(Model):
             # Raise if the request failed because of a malformed function call
             if hasattr(candidate, "finish_reason") and candidate.finish_reason:
                 if candidate.finish_reason == GeminiFinishReason.MALFORMED_FUNCTION_CALL.value:
-                    # We only want to raise errors that trigger regeneration attempts once
-                    if kwargs.get("retrying_with_guidance") is True:
-                        pass
                     if self.retry_with_guidance:
-                        raise RetryableModelProviderError(retry_guidance_message=MALFORMED_FUNCTION_CALL_GUIDANCE)
+                        raise RetryableModelProviderError(
+                            retry_guidance_message=MALFORMED_FUNCTION_CALL_GUIDANCE,
+                            original_error=f"Generation ended with finish reason: {candidate.finish_reason}",
+                        )
 
             if candidate.content:
                 response_message = candidate.content
@@ -1079,9 +1195,11 @@ class Gemini(Model):
             # Raise if the request failed because of a malformed function call
             if hasattr(candidate, "finish_reason") and candidate.finish_reason:
                 if candidate.finish_reason == GeminiFinishReason.MALFORMED_FUNCTION_CALL.value:
-                    if kwargs.get("retrying_with_guidance") is True:
-                        pass
-                    raise RetryableModelProviderError(retry_guidance_message=MALFORMED_FUNCTION_CALL_GUIDANCE)
+                    if self.retry_with_guidance:
+                        raise RetryableModelProviderError(
+                            retry_guidance_message=MALFORMED_FUNCTION_CALL_GUIDANCE,
+                            original_error=f"Generation ended with finish reason: {candidate.finish_reason}",
+                        )
 
             response_message: Content = Content(role="model", parts=[])
             if candidate_content is not None:

agno/models/litellm/chat.py

@@ -10,8 +10,10 @@ from agno.models.message import Message
 from agno.models.metrics import Metrics
 from agno.models.response import ModelResponse
 from agno.run.agent import RunOutput
+from agno.tools.function import Function
 from agno.utils.log import log_debug, log_error, log_warning
 from agno.utils.openai import _format_file_for_message, audio_to_message, images_to_message
+from agno.utils.tokens import count_schema_tokens
 
 try:
     import litellm
@@ -476,3 +478,26 @@ class LiteLLM(Model):
         metrics.total_tokens = metrics.input_tokens + metrics.output_tokens
 
         return metrics
+
+    def count_tokens(
+        self,
+        messages: List[Message],
+        tools: Optional[List[Union[Function, Dict[str, Any]]]] = None,
+        response_format: Optional[Union[Dict, Type[BaseModel]]] = None,
+    ) -> int:
+        formatted_messages = self._format_messages(messages, compress_tool_results=True)
+        formatted_tools = self._format_tools(tools) if tools else None
+        tokens = litellm.token_counter(
+            model=self.id,
+            messages=formatted_messages,
+            tools=formatted_tools,  # type: ignore
+        )
+        return tokens + count_schema_tokens(response_format, self.id)
+
+    async def acount_tokens(
+        self,
+        messages: List[Message],
+        tools: Optional[List[Union[Function, Dict[str, Any]]]] = None,
+        response_format: Optional[Union[Dict, Type[BaseModel]]] = None,
+    ) -> int:
+        return self.count_tokens(messages, tools, response_format)