Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

feat: add structured outputs and response_format support across providers (#4280)

Greyson LaLonde97975673424e
diff --git a/lib/crewai/pyproject.toml b/lib/crewai/pyproject.toml
index bd01dbf9a..4a21e1fa6 100644
--- a/lib/crewai/pyproject.toml
+++ b/lib/crewai/pyproject.toml
@@ -90,7 +90,7 @@ azure-ai-inference = [
"azure-ai-inference~=1.0.0b9",
]
anthropic = [
- "anthropic~=0.71.0",
+ "anthropic~=0.73.0",
]
a2a = [
"a2a-sdk~=0.3.10",
diff --git a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
index 1d626a7f4..e98249567 100644
--- a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
@@ -3,9 +3,8 @@ from __future__ import annotations
import json
import logging
import os
-from typing import TYPE_CHECKING, Any, Literal, cast
+from typing import TYPE_CHECKING, Any, Final, Literal, TypeGuard, cast
-from anthropic.types import ThinkingBlock
from pydantic import BaseModel
from crewai.events.types.llm_events import LLMCallType
@@ -22,8 +21,9 @@ if TYPE_CHECKING:
from crewai.llms.hooks.base import BaseInterceptor
try:
- from anthropic import Anthropic, AsyncAnthropic
+ from anthropic import Anthropic, AsyncAnthropic, transform_schema
from anthropic.types import Message, TextBlock, ThinkingBlock, ToolUseBlock
+ from anthropic.types.beta import BetaMessage
import httpx
except ImportError:
raise ImportError(
@@ -31,7 +31,62 @@ except ImportError:
) from None
-ANTHROPIC_FILES_API_BETA = "files-api-2025-04-14"
+ANTHROPIC_FILES_API_BETA: Final = "files-api-2025-04-14"
+ANTHROPIC_STRUCTURED_OUTPUTS_BETA: Final = "structured-outputs-2025-11-13"
+
+NATIVE_STRUCTURED_OUTPUT_MODELS: Final[
+ tuple[
+ Literal["claude-sonnet-4-5"],
+ Literal["claude-sonnet-4.5"],
+ Literal["claude-opus-4-5"],
+ Literal["claude-opus-4.5"],
+ Literal["claude-opus-4-1"],
+ Literal["claude-opus-4.1"],
+ Literal["claude-haiku-4-5"],
+ Literal["claude-haiku-4.5"],
+ ]
+] = (
+ "claude-sonnet-4-5",
+ "claude-sonnet-4.5",
+ "claude-opus-4-5",
+ "claude-opus-4.5",
+ "claude-opus-4-1",
+ "claude-opus-4.1",
+ "claude-haiku-4-5",
+ "claude-haiku-4.5",
+)
+
+
+def _supports_native_structured_outputs(model: str) -> bool:
+ """Check if the model supports native structured outputs.
+
+ Native structured outputs are only available for Claude 4.5 models
+ (Sonnet 4.5, Opus 4.5, Opus 4.1, Haiku 4.5).
+ Other models require the tool-based fallback approach.
+
+ Args:
+ model: The model name/identifier.
+
+ Returns:
+ True if the model supports native structured outputs.
+ """
+ model_lower = model.lower()
+ return any(prefix in model_lower for prefix in NATIVE_STRUCTURED_OUTPUT_MODELS)
+
+
+def _is_pydantic_model_class(obj: Any) -> TypeGuard[type[BaseModel]]:
+ """Check if an object is a Pydantic model class.
+
+ This distinguishes between Pydantic model classes that support structured
+ outputs (have model_json_schema) and plain dicts like {"type": "json_object"}.
+
+ Args:
+ obj: The object to check.
+
+ Returns:
+ True if obj is a Pydantic model class.
+ """
+ return isinstance(obj, type) and issubclass(obj, BaseModel)
def _contains_file_id_reference(messages: list[dict[str, Any]]) -> bool:
@@ -84,6 +139,7 @@ class AnthropicCompletion(BaseLLM):
client_params: dict[str, Any] | None = None,
interceptor: BaseInterceptor[httpx.Request, httpx.Response] | None = None,
thinking: AnthropicThinkingConfig | None = None,
+ response_format: type[BaseModel] | None = None,
**kwargs: Any,
):
"""Initialize Anthropic chat completion client.
@@ -101,6 +157,8 @@ class AnthropicCompletion(BaseLLM):
stream: Enable streaming responses
client_params: Additional parameters for the Anthropic client
interceptor: HTTP interceptor for modifying requests/responses at transport level.
+ response_format: Pydantic model for structured output. When provided, responses
+ will be validated against this model schema.
**kwargs: Additional parameters
"""
super().__init__(
@@ -131,6 +189,7 @@ class AnthropicCompletion(BaseLLM):
self.stop_sequences = stop_sequences or []
self.thinking = thinking
self.previous_thinking_blocks: list[ThinkingBlock] = []
+ self.response_format = response_format
# Model-specific settings
self.is_claude_3 = "claude-3" in model.lower()
self.supports_tools = True
@@ -231,6 +290,8 @@ class AnthropicCompletion(BaseLLM):
formatted_messages, system_message, tools
)
+ effective_response_model = response_model or self.response_format
+
# Handle streaming vs non-streaming
if self.stream:
return self._handle_streaming_completion(
@@ -238,7 +299,7 @@ class AnthropicCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
return self._handle_completion(
@@ -246,7 +307,7 @@ class AnthropicCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
except Exception as e:
@@ -298,13 +359,15 @@ class AnthropicCompletion(BaseLLM):
formatted_messages, system_message, tools
)
+ effective_response_model = response_model or self.response_format
+
if self.stream:
return await self._ahandle_streaming_completion(
completion_params,
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
return await self._ahandle_completion(
@@ -312,7 +375,7 @@ class AnthropicCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
except Exception as e:
@@ -565,22 +628,40 @@ class AnthropicCompletion(BaseLLM):
response_model: type[BaseModel] | None = None,
) -> str | Any:
"""Handle non-streaming message completion."""
- if response_model:
- structured_tool = {
- "name": "structured_output",
- "description": "Returns structured data according to the schema",
- "input_schema": response_model.model_json_schema(),
- }
-
- params["tools"] = [structured_tool]
- params["tool_choice"] = {"type": "tool", "name": "structured_output"}
-
uses_file_api = _contains_file_id_reference(params.get("messages", []))
+ betas: list[str] = []
+ use_native_structured_output = False
+
+ if uses_file_api:
+ betas.append(ANTHROPIC_FILES_API_BETA)
+
+ extra_body: dict[str, Any] | None = None
+ if _is_pydantic_model_class(response_model):
+ schema = transform_schema(response_model.model_json_schema())
+ if _supports_native_structured_outputs(self.model):
+ use_native_structured_output = True
+ betas.append(ANTHROPIC_STRUCTURED_OUTPUTS_BETA)
+ extra_body = {
+ "output_format": {
+ "type": "json_schema",
+ "schema": schema,
+ }
+ }
+ else:
+ structured_tool = {
+ "name": "structured_output",
+ "description": "Output the structured response",
+ "input_schema": schema,
+ }
+ params["tools"] = [structured_tool]
+ params["tool_choice"] = {"type": "tool", "name": "structured_output"}
try:
- if uses_file_api:
- params["betas"] = [ANTHROPIC_FILES_API_BETA]
- response = self.client.beta.messages.create(**params)
+ if betas:
+ params["betas"] = betas
+ response = self.client.beta.messages.create(
+ **params, extra_body=extra_body
+ )
else:
response = self.client.messages.create(**params)
@@ -593,22 +674,34 @@ class AnthropicCompletion(BaseLLM):
usage = self._extract_anthropic_token_usage(response)
self._track_token_usage_internal(usage)
- if response_model and response.content:
- tool_uses = [
- block for block in response.content if isinstance(block, ToolUseBlock)
- ]
- if tool_uses and tool_uses[0].name == "structured_output":
- structured_data = tool_uses[0].input
- structured_json = json.dumps(structured_data)
- self._emit_call_completed_event(
- response=structured_json,
- call_type=LLMCallType.LLM_CALL,
- from_task=from_task,
- from_agent=from_agent,
- messages=params["messages"],
- )
-
- return structured_json
+ if _is_pydantic_model_class(response_model) and response.content:
+ if use_native_structured_output:
+ for block in response.content:
+ if isinstance(block, TextBlock):
+ structured_json = block.text
+ self._emit_call_completed_event(
+ response=structured_json,
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=params["messages"],
+ )
+ return structured_json
+ else:
+ for block in response.content:
+ if (
+ isinstance(block, ToolUseBlock)
+ and block.name == "structured_output"
+ ):
+ structured_json = json.dumps(block.input)
+ self._emit_call_completed_event(
+ response=structured_json,
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=params["messages"],
+ )
+ return structured_json
# Check if Claude wants to use tools
if response.content:
@@ -678,17 +771,31 @@ class AnthropicCompletion(BaseLLM):
from_task: Any | None = None,
from_agent: Any | None = None,
response_model: type[BaseModel] | None = None,
- ) -> str:
+ ) -> str | Any:
"""Handle streaming message completion."""
- if response_model:
- structured_tool = {
- "name": "structured_output",
- "description": "Returns structured data according to the schema",
- "input_schema": response_model.model_json_schema(),
- }
-
- params["tools"] = [structured_tool]
- params["tool_choice"] = {"type": "tool", "name": "structured_output"}
+ betas: list[str] = []
+ use_native_structured_output = False
+
+ extra_body: dict[str, Any] | None = None
+ if _is_pydantic_model_class(response_model):
+ schema = transform_schema(response_model.model_json_schema())
+ if _supports_native_structured_outputs(self.model):
+ use_native_structured_output = True
+ betas.append(ANTHROPIC_STRUCTURED_OUTPUTS_BETA)
+ extra_body = {
+ "output_format": {
+ "type": "json_schema",
+ "schema": schema,
+ }
+ }
+ else:
+ structured_tool = {
+ "name": "structured_output",
+ "description": "Output the structured response",
+ "input_schema": schema,
+ }
+ params["tools"] = [structured_tool]
+ params["tool_choice"] = {"type": "tool", "name": "structured_output"}
full_response = ""
@@ -696,15 +803,22 @@ class AnthropicCompletion(BaseLLM):
# (the SDK sets it internally)
stream_params = {k: v for k, v in params.items() if k != "stream"}
+ if betas:
+ stream_params["betas"] = betas
+
current_tool_calls: dict[int, dict[str, Any]] = {}
- # Make streaming API call
- with self.client.messages.stream(**stream_params) as stream:
+ stream_context = (
+ self.client.beta.messages.stream(**stream_params, extra_body=extra_body)
+ if betas
+ else self.client.messages.stream(**stream_params)
+ )
+ with stream_context as stream:
response_id = None
for event in stream:
if hasattr(event, "message") and hasattr(event.message, "id"):
response_id = event.message.id
-
+
if hasattr(event, "delta") and hasattr(event.delta, "text"):
text_delta = event.delta.text
full_response += text_delta
@@ -712,7 +826,7 @@ class AnthropicCompletion(BaseLLM):
chunk=text_delta,
from_task=from_task,
from_agent=from_agent,
- response_id=response_id
+ response_id=response_id,
)
if event.type == "content_block_start":
@@ -739,7 +853,7 @@ class AnthropicCompletion(BaseLLM):
"index": block_index,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
elif event.type == "content_block_delta":
if event.delta.type == "input_json_delta":
@@ -763,10 +877,10 @@ class AnthropicCompletion(BaseLLM):
"index": block_index,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
- final_message: Message = stream.get_final_message()
+ final_message = stream.get_final_message()
thinking_blocks: list[ThinkingBlock] = []
if final_message.content:
@@ -781,25 +895,30 @@ class AnthropicCompletion(BaseLLM):
usage = self._extract_anthropic_token_usage(final_message)
self._track_token_usage_internal(usage)
- if response_model and final_message.content:
- tool_uses = [
- block
- for block in final_message.content
- if isinstance(block, ToolUseBlock)
- ]
- if tool_uses and tool_uses[0].name == "structured_output":
- structured_data = tool_uses[0].input
- structured_json = json.dumps(structured_data)
-
+ if _is_pydantic_model_class(response_model):
+ if use_native_structured_output:
self._emit_call_completed_event(
- response=structured_json,
+ response=full_response,
call_type=LLMCallType.LLM_CALL,
from_task=from_task,
from_agent=from_agent,
messages=params["messages"],
)
-
- return structured_json
+ return full_response
+ for block in final_message.content:
+ if (
+ isinstance(block, ToolUseBlock)
+ and block.name == "structured_output"
+ ):
+ structured_json = json.dumps(block.input)
+ self._emit_call_completed_event(
+ response=structured_json,
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=params["messages"],
+ )
+ return structured_json
if final_message.content:
tool_uses = [
@@ -809,11 +928,9 @@ class AnthropicCompletion(BaseLLM):
]
if tool_uses:
- # If no available_functions, return tool calls for executor to handle
if not available_functions:
return list(tool_uses)
- # Handle tool use conversation flow internally
return self._handle_tool_use_conversation(
final_message,
tool_uses,
@@ -823,10 +940,8 @@ class AnthropicCompletion(BaseLLM):
from_agent,
)
- # Apply stop words to full response
full_response = self._apply_stop_words(full_response)
- # Emit completion event and return full response
self._emit_call_completed_event(
response=full_response,
call_type=LLMCallType.LLM_CALL,
@@ -884,7 +999,7 @@ class AnthropicCompletion(BaseLLM):
def _handle_tool_use_conversation(
self,
- initial_response: Message,
+ initial_response: Message | BetaMessage,
tool_uses: list[ToolUseBlock],
params: dict[str, Any],
available_functions: dict[str, Any],
@@ -1002,22 +1117,40 @@ class AnthropicCompletion(BaseLLM):
response_model: type[BaseModel] | None = None,
) -> str | Any:
"""Handle non-streaming async message completion."""
- if response_model:
- structured_tool = {
- "name": "structured_output",
- "description": "Returns structured data according to the schema",
- "input_schema": response_model.model_json_schema(),
- }
-
- params["tools"] = [structured_tool]
- params["tool_choice"] = {"type": "tool", "name": "structured_output"}
-
uses_file_api = _contains_file_id_reference(params.get("messages", []))
+ betas: list[str] = []
+ use_native_structured_output = False
+
+ if uses_file_api:
+ betas.append(ANTHROPIC_FILES_API_BETA)
+
+ extra_body: dict[str, Any] | None = None
+ if _is_pydantic_model_class(response_model):
+ schema = transform_schema(response_model.model_json_schema())
+ if _supports_native_structured_outputs(self.model):
+ use_native_structured_output = True
+ betas.append(ANTHROPIC_STRUCTURED_OUTPUTS_BETA)
+ extra_body = {
+ "output_format": {
+ "type": "json_schema",
+ "schema": schema,
+ }
+ }
+ else:
+ structured_tool = {
+ "name": "structured_output",
+ "description": "Output the structured response",
+ "input_schema": schema,
+ }
+ params["tools"] = [structured_tool]
+ params["tool_choice"] = {"type": "tool", "name": "structured_output"}
try:
- if uses_file_api:
- params["betas"] = [ANTHROPIC_FILES_API_BETA]
- response = await self.async_client.beta.messages.create(**params)
+ if betas:
+ params["betas"] = betas
+ response = await self.async_client.beta.messages.create(
+ **params, extra_body=extra_body
+ )
else:
response = await self.async_client.messages.create(**params)
@@ -1030,23 +1163,34 @@ class AnthropicCompletion(BaseLLM):
usage = self._extract_anthropic_token_usage(response)
self._track_token_usage_internal(usage)
- if response_model and response.content:
- tool_uses = [
- block for block in response.content if isinstance(block, ToolUseBlock)
- ]
- if tool_uses and tool_uses[0].name == "structured_output":
- structured_data = tool_uses[0].input
- structured_json = json.dumps(structured_data)
-
- self._emit_call_completed_event(
- response=structured_json,
- call_type=LLMCallType.LLM_CALL,
- from_task=from_task,
- from_agent=from_agent,
- messages=params["messages"],
- )
-
- return structured_json
+ if _is_pydantic_model_class(response_model) and response.content:
+ if use_native_structured_output:
+ for block in response.content:
+ if isinstance(block, TextBlock):
+ structured_json = block.text
+ self._emit_call_completed_event(
+ response=structured_json,
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=params["messages"],
+ )
+ return structured_json
+ else:
+ for block in response.content:
+ if (
+ isinstance(block, ToolUseBlock)
+ and block.name == "structured_output"
+ ):
+ structured_json = json.dumps(block.input)
+ self._emit_call_completed_event(
+ response=structured_json,
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=params["messages"],
+ )
+ return structured_json
if response.content:
tool_uses = [
@@ -1102,25 +1246,49 @@ class AnthropicCompletion(BaseLLM):
from_task: Any | None = None,
from_agent: Any | None = None,
response_model: type[BaseModel] | None = None,
- ) -> str:
+ ) -> str | Any:
"""Handle async streaming message completion."""
- if response_model:
- structured_tool = {
- "name": "structured_output",
- "description": "Returns structured data according to the schema",
- "input_schema": response_model.model_json_schema(),
- }
-
- params["tools"] = [structured_tool]
- params["tool_choice"] = {"type": "tool", "name": "structured_output"}
+ betas: list[str] = []
+ use_native_structured_output = False
+
+ extra_body: dict[str, Any] | None = None
+ if _is_pydantic_model_class(response_model):
+ schema = transform_schema(response_model.model_json_schema())
+ if _supports_native_structured_outputs(self.model):
+ use_native_structured_output = True
+ betas.append(ANTHROPIC_STRUCTURED_OUTPUTS_BETA)
+ extra_body = {
+ "output_format": {
+ "type": "json_schema",
+ "schema": schema,
+ }
+ }
+ else:
+ structured_tool = {
+ "name": "structured_output",
+ "description": "Output the structured response",
+ "input_schema": schema,
+ }
+ params["tools"] = [structured_tool]
+ params["tool_choice"] = {"type": "tool", "name": "structured_output"}
full_response = ""
stream_params = {k: v for k, v in params.items() if k != "stream"}
+ if betas:
+ stream_params["betas"] = betas
+
current_tool_calls: dict[int, dict[str, Any]] = {}
- async with self.async_client.messages.stream(**stream_params) as stream:
+ stream_context = (
+ self.async_client.beta.messages.stream(
+ **stream_params, extra_body=extra_body
+ )
+ if betas
+ else self.async_client.messages.stream(**stream_params)
+ )
+ async with stream_context as stream:
response_id = None
async for event in stream:
if hasattr(event, "message") and hasattr(event.message, "id"):
@@ -1133,7 +1301,7 @@ class AnthropicCompletion(BaseLLM):
chunk=text_delta,
from_task=from_task,
from_agent=from_agent,
- response_id=response_id
+ response_id=response_id,
)
if event.type == "content_block_start":
@@ -1160,7 +1328,7 @@ class AnthropicCompletion(BaseLLM):
"index": block_index,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
elif event.type == "content_block_delta":
if event.delta.type == "input_json_delta":
@@ -1184,33 +1352,38 @@ class AnthropicCompletion(BaseLLM):
"index": block_index,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
- final_message: Message = await stream.get_final_message()
+ final_message = await stream.get_final_message()
usage = self._extract_anthropic_token_usage(final_message)
self._track_token_usage_internal(usage)
- if response_model and final_message.content:
- tool_uses = [
- block
- for block in final_message.content
- if isinstance(block, ToolUseBlock)
- ]
- if tool_uses and tool_uses[0].name == "structured_output":
- structured_data = tool_uses[0].input
- structured_json = json.dumps(structured_data)
-
+ if _is_pydantic_model_class(response_model):
+ if use_native_structured_output:
self._emit_call_completed_event(
- response=structured_json,
+ response=full_response,
call_type=LLMCallType.LLM_CALL,
from_task=from_task,
from_agent=from_agent,
messages=params["messages"],
)
-
- return structured_json
+ return full_response
+ for block in final_message.content:
+ if (
+ isinstance(block, ToolUseBlock)
+ and block.name == "structured_output"
+ ):
+ structured_json = json.dumps(block.input)
+ self._emit_call_completed_event(
+ response=structured_json,
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=params["messages"],
+ )
+ return structured_json
if final_message.content:
tool_uses = [
@@ -1220,7 +1393,6 @@ class AnthropicCompletion(BaseLLM):
]
if tool_uses:
- # If no available_functions, return tool calls for executor to handle
if not available_functions:
return list(tool_uses)
@@ -1247,7 +1419,7 @@ class AnthropicCompletion(BaseLLM):
async def _ahandle_tool_use_conversation(
self,
- initial_response: Message,
+ initial_response: Message | BetaMessage,
tool_uses: list[ToolUseBlock],
params: dict[str, Any],
available_functions: dict[str, Any],
@@ -1356,7 +1528,9 @@ class AnthropicCompletion(BaseLLM):
return int(200000 * CONTEXT_WINDOW_USAGE_RATIO)
@staticmethod
- def _extract_anthropic_token_usage(response: Message) -> dict[str, Any]:
+ def _extract_anthropic_token_usage(
+ response: Message | BetaMessage,
+ ) -> dict[str, Any]:
"""Extract token usage from Anthropic response."""
if hasattr(response, "usage") and response.usage:
usage = response.usage
diff --git a/lib/crewai/src/crewai/llms/providers/azure/completion.py b/lib/crewai/src/crewai/llms/providers/azure/completion.py
index a3aed7f4b..1de18d984 100644
--- a/lib/crewai/src/crewai/llms/providers/azure/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/azure/completion.py
@@ -92,6 +92,7 @@ class AzureCompletion(BaseLLM):
stop: list[str] | None = None,
stream: bool = False,
interceptor: BaseInterceptor[Any, Any] | None = None,
+ response_format: type[BaseModel] | None = None,
**kwargs: Any,
):
"""Initialize Azure AI Inference chat completion client.
@@ -111,6 +112,9 @@ class AzureCompletion(BaseLLM):
stop: Stop sequences
stream: Enable streaming responses
interceptor: HTTP interceptor (not yet supported for Azure).
+ response_format: Pydantic model for structured output. Used as default when
+ response_model is not passed to call()/acall() methods.
+ Only works with OpenAI models deployed on Azure.
**kwargs: Additional parameters
"""
if interceptor is not None:
@@ -165,6 +169,7 @@ class AzureCompletion(BaseLLM):
self.presence_penalty = presence_penalty
self.max_tokens = max_tokens
self.stream = stream
+ self.response_format = response_format
self.is_openai_model = any(
prefix in model.lower() for prefix in ["gpt-", "o1-", "text-"]
@@ -298,6 +303,7 @@ class AzureCompletion(BaseLLM):
from_task=from_task,
from_agent=from_agent,
)
+ effective_response_model = response_model or self.response_format
# Format messages for Azure
formatted_messages = self._format_messages_for_azure(messages)
@@ -307,7 +313,7 @@ class AzureCompletion(BaseLLM):
# Prepare completion parameters
completion_params = self._prepare_completion_params(
- formatted_messages, tools, response_model
+ formatted_messages, tools, effective_response_model
)
# Handle streaming vs non-streaming
@@ -317,7 +323,7 @@ class AzureCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
return self._handle_completion(
@@ -325,7 +331,7 @@ class AzureCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
except Exception as e:
@@ -364,11 +370,12 @@ class AzureCompletion(BaseLLM):
from_task=from_task,
from_agent=from_agent,
)
+ effective_response_model = response_model or self.response_format
formatted_messages = self._format_messages_for_azure(messages)
completion_params = self._prepare_completion_params(
- formatted_messages, tools, response_model
+ formatted_messages, tools, effective_response_model
)
if self.stream:
@@ -377,7 +384,7 @@ class AzureCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
return await self._ahandle_completion(
@@ -385,7 +392,7 @@ class AzureCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
except Exception as e:
@@ -726,7 +733,7 @@ class AzureCompletion(BaseLLM):
"""
if update.choices:
choice = update.choices[0]
- response_id = update.id if hasattr(update,"id") else None
+ response_id = update.id if hasattr(update, "id") else None
if choice.delta and choice.delta.content:
content_delta = choice.delta.content
full_response += content_delta
@@ -734,7 +741,7 @@ class AzureCompletion(BaseLLM):
chunk=content_delta,
from_task=from_task,
from_agent=from_agent,
- response_id=response_id
+ response_id=response_id,
)
if choice.delta and choice.delta.tool_calls:
@@ -769,7 +776,7 @@ class AzureCompletion(BaseLLM):
"index": idx,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
return full_response
diff --git a/lib/crewai/src/crewai/llms/providers/bedrock/completion.py b/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
index 73fefc478..2ab638f11 100644
--- a/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
@@ -172,6 +172,7 @@ class BedrockCompletion(BaseLLM):
additional_model_request_fields: dict[str, Any] | None = None,
additional_model_response_field_paths: list[str] | None = None,
interceptor: BaseInterceptor[Any, Any] | None = None,
+ response_format: type[BaseModel] | None = None,
**kwargs: Any,
) -> None:
"""Initialize AWS Bedrock completion client.
@@ -192,6 +193,8 @@ class BedrockCompletion(BaseLLM):
additional_model_request_fields: Model-specific request parameters
additional_model_response_field_paths: Custom response field paths
interceptor: HTTP interceptor (not yet supported for Bedrock).
+ response_format: Pydantic model for structured output. Used as default when
+ response_model is not passed to call()/acall() methods.
**kwargs: Additional parameters
"""
if interceptor is not None:
@@ -248,6 +251,7 @@ class BedrockCompletion(BaseLLM):
self.top_k = top_k
self.stream = stream
self.stop_sequences = stop_sequences
+ self.response_format = response_format
# Store advanced features (optional)
self.guardrail_config = guardrail_config
@@ -299,6 +303,8 @@ class BedrockCompletion(BaseLLM):
response_model: type[BaseModel] | None = None,
) -> str | Any:
"""Call AWS Bedrock Converse API."""
+ effective_response_model = response_model or self.response_format
+
try:
# Emit call started event
self._emit_call_started_event(
@@ -375,6 +381,7 @@ class BedrockCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
+ effective_response_model,
)
return self._handle_converse(
@@ -383,6 +390,7 @@ class BedrockCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
+ effective_response_model,
)
except Exception as e:
@@ -425,6 +433,8 @@ class BedrockCompletion(BaseLLM):
NotImplementedError: If aiobotocore is not installed.
LLMContextLengthExceededError: If context window is exceeded.
"""
+ effective_response_model = response_model or self.response_format
+
if not AIOBOTOCORE_AVAILABLE:
raise NotImplementedError(
"Async support for AWS Bedrock requires aiobotocore. "
@@ -494,11 +504,21 @@ class BedrockCompletion(BaseLLM):
if self.stream:
return await self._ahandle_streaming_converse(
- formatted_messages, body, available_functions, from_task, from_agent
+ formatted_messages,
+ body,
+ available_functions,
+ from_task,
+ from_agent,
+ effective_response_model,
)
return await self._ahandle_converse(
- formatted_messages, body, available_functions, from_task, from_agent
+ formatted_messages,
+ body,
+ available_functions,
+ from_task,
+ from_agent,
+ effective_response_model,
)
except Exception as e:
@@ -520,10 +540,29 @@ class BedrockCompletion(BaseLLM):
available_functions: Mapping[str, Any] | None = None,
from_task: Any | None = None,
from_agent: Any | None = None,
- ) -> str:
+ response_model: type[BaseModel] | None = None,
+ ) -> str | Any:
"""Handle non-streaming converse API call following AWS best practices."""
+ if response_model:
+ structured_tool: ConverseToolTypeDef = {
+ "toolSpec": {
+ "name": "structured_output",
+ "description": "Returns structured data according to the schema",
+ "inputSchema": {"json": response_model.model_json_schema()},
+ }
+ }
+ body["toolConfig"] = cast(
+ "ToolConfigurationTypeDef",
+ cast(
+ object,
+ {
+ "tools": [structured_tool],
+ "toolChoice": {"tool": {"name": "structured_output"}},
+ },
+ ),
+ )
+
try:
- # Validate messages format before API call
if not messages:
raise ValueError("Messages cannot be empty")
@@ -571,6 +610,21 @@ class BedrockCompletion(BaseLLM):
# If there are tool uses but no available_functions, return them for the executor to handle
tool_uses = [block["toolUse"] for block in content if "toolUse" in block]
+
+ if response_model and tool_uses:
+ for tool_use in tool_uses:
+ if tool_use.get("name") == "structured_output":
+ structured_data = tool_use.get("input", {})
+ result = response_model.model_validate(structured_data)
+ self._emit_call_completed_event(
+ response=result.model_dump_json(),
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=messages,
+ )
+ return result
+
if tool_uses and not available_functions:
self._emit_call_completed_event(
response=tool_uses,
@@ -717,8 +771,28 @@ class BedrockCompletion(BaseLLM):
available_functions: dict[str, Any] | None = None,
from_task: Any | None = None,
from_agent: Any | None = None,
+ response_model: type[BaseModel] | None = None,
) -> str:
"""Handle streaming converse API call with comprehensive event handling."""
+ if response_model:
+ structured_tool: ConverseToolTypeDef = {
+ "toolSpec": {
+ "name": "structured_output",
+ "description": "Returns structured data according to the schema",
+ "inputSchema": {"json": response_model.model_json_schema()},
+ }
+ }
+ body["toolConfig"] = cast(
+ "ToolConfigurationTypeDef",
+ cast(
+ object,
+ {
+ "tools": [structured_tool],
+ "toolChoice": {"tool": {"name": "structured_output"}},
+ },
+ ),
+ )
+
full_response = ""
current_tool_use: dict[str, Any] | None = None
tool_use_id: str | None = None
@@ -805,7 +879,7 @@ class BedrockCompletion(BaseLLM):
"index": tool_use_index,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
elif "contentBlockStop" in event:
logging.debug("Content block stopped in stream")
@@ -929,8 +1003,28 @@ class BedrockCompletion(BaseLLM):
available_functions: Mapping[str, Any] | None = None,
from_task: Any | None = None,
from_agent: Any | None = None,
- ) -> str:
+ response_model: type[BaseModel] | None = None,
+ ) -> str | Any:
"""Handle async non-streaming converse API call."""
+ if response_model:
+ structured_tool: ConverseToolTypeDef = {
+ "toolSpec": {
+ "name": "structured_output",
+ "description": "Returns structured data according to the schema",
+ "inputSchema": {"json": response_model.model_json_schema()},
+ }
+ }
+ body["toolConfig"] = cast(
+ "ToolConfigurationTypeDef",
+ cast(
+ object,
+ {
+ "tools": [structured_tool],
+ "toolChoice": {"tool": {"name": "structured_output"}},
+ },
+ ),
+ )
+
try:
if not messages:
raise ValueError("Messages cannot be empty")
@@ -976,6 +1070,21 @@ class BedrockCompletion(BaseLLM):
# If there are tool uses but no available_functions, return them for the executor to handle
tool_uses = [block["toolUse"] for block in content if "toolUse" in block]
+
+ if response_model and tool_uses:
+ for tool_use in tool_uses:
+ if tool_use.get("name") == "structured_output":
+ structured_data = tool_use.get("input", {})
+ result = response_model.model_validate(structured_data)
+ self._emit_call_completed_event(
+ response=result.model_dump_json(),
+ call_type=LLMCallType.LLM_CALL,
+ from_task=from_task,
+ from_agent=from_agent,
+ messages=messages,
+ )
+ return result
+
if tool_uses and not available_functions:
self._emit_call_completed_event(
response=tool_uses,
@@ -1106,8 +1215,28 @@ class BedrockCompletion(BaseLLM):
available_functions: dict[str, Any] | None = None,
from_task: Any | None = None,
from_agent: Any | None = None,
+ response_model: type[BaseModel] | None = None,
) -> str:
"""Handle async streaming converse API call."""
+ if response_model:
+ structured_tool: ConverseToolTypeDef = {
+ "toolSpec": {
+ "name": "structured_output",
+ "description": "Returns structured data according to the schema",
+ "inputSchema": {"json": response_model.model_json_schema()},
+ }
+ }
+ body["toolConfig"] = cast(
+ "ToolConfigurationTypeDef",
+ cast(
+ object,
+ {
+ "tools": [structured_tool],
+ "toolChoice": {"tool": {"name": "structured_output"}},
+ },
+ ),
+ )
+
full_response = ""
current_tool_use: dict[str, Any] | None = None
tool_use_id: str | None = None
@@ -1174,7 +1303,7 @@ class BedrockCompletion(BaseLLM):
chunk=text_chunk,
from_task=from_task,
from_agent=from_agent,
- response_id=response_id
+ response_id=response_id,
)
elif "toolUse" in delta and current_tool_use:
tool_input = delta["toolUse"].get("input", "")
diff --git a/lib/crewai/src/crewai/llms/providers/gemini/completion.py b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
index 9687f3d4f..d101ad0be 100644
--- a/lib/crewai/src/crewai/llms/providers/gemini/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
@@ -56,6 +56,7 @@ class GeminiCompletion(BaseLLM):
client_params: dict[str, Any] | None = None,
interceptor: BaseInterceptor[Any, Any] | None = None,
use_vertexai: bool | None = None,
+ response_format: type[BaseModel] | None = None,
**kwargs: Any,
):
"""Initialize Google Gemini chat completion client.
@@ -86,6 +87,8 @@ class GeminiCompletion(BaseLLM):
- None (default): Check GOOGLE_GENAI_USE_VERTEXAI env var
When using Vertex AI with API key (Express mode), http_options with
api_version="v1" is automatically configured.
+ response_format: Pydantic model for structured output. Used as default when
+ response_model is not passed to call()/acall() methods.
**kwargs: Additional parameters
"""
if interceptor is not None:
@@ -121,6 +124,7 @@ class GeminiCompletion(BaseLLM):
self.safety_settings = safety_settings or {}
self.stop_sequences = stop_sequences or []
self.tools: list[dict[str, Any]] | None = None
+ self.response_format = response_format
# Model-specific settings
version_match = re.search(r"gemini-(\d+(?:\.\d+)?)", model.lower())
@@ -292,6 +296,7 @@ class GeminiCompletion(BaseLLM):
from_agent=from_agent,
)
self.tools = tools
+ effective_response_model = response_model or self.response_format
formatted_content, system_instruction = self._format_messages_for_gemini(
messages
@@ -303,7 +308,7 @@ class GeminiCompletion(BaseLLM):
raise ValueError("LLM call blocked by before_llm_call hook")
config = self._prepare_generation_config(
- system_instruction, tools, response_model
+ system_instruction, tools, effective_response_model
)
if self.stream:
@@ -313,7 +318,7 @@ class GeminiCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
return self._handle_completion(
@@ -322,7 +327,7 @@ class GeminiCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
except APIError as e:
@@ -374,13 +379,14 @@ class GeminiCompletion(BaseLLM):
from_agent=from_agent,
)
self.tools = tools
+ effective_response_model = response_model or self.response_format
formatted_content, system_instruction = self._format_messages_for_gemini(
messages
)
config = self._prepare_generation_config(
- system_instruction, tools, response_model
+ system_instruction, tools, effective_response_model
)
if self.stream:
@@ -390,7 +396,7 @@ class GeminiCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
return await self._ahandle_completion(
@@ -399,7 +405,7 @@ class GeminiCompletion(BaseLLM):
available_functions,
from_task,
from_agent,
- response_model,
+ effective_response_model,
)
except APIError as e:
@@ -570,10 +576,10 @@ class GeminiCompletion(BaseLLM):
types.Content(role="user", parts=[function_response_part])
)
elif role == "assistant" and message.get("tool_calls"):
- parts: list[types.Part] = []
+ tool_parts: list[types.Part] = []
if text_content:
- parts.append(types.Part.from_text(text=text_content))
+ tool_parts.append(types.Part.from_text(text=text_content))
tool_calls: list[dict[str, Any]] = message.get("tool_calls") or []
for tool_call in tool_calls:
@@ -592,11 +598,11 @@ class GeminiCompletion(BaseLLM):
else:
func_args = func_args_raw
- parts.append(
+ tool_parts.append(
types.Part.from_function_call(name=func_name, args=func_args)
)
- contents.append(types.Content(role="model", parts=parts))
+ contents.append(types.Content(role="model", parts=tool_parts))
else:
# Convert role for Gemini (assistant -> model)
gemini_role = "model" if role == "assistant" else "user"
@@ -790,7 +796,7 @@ class GeminiCompletion(BaseLLM):
Returns:
Tuple of (updated full_response, updated function_calls, updated usage_data)
"""
- response_id=chunk.response_id if hasattr(chunk,"response_id") else None
+ response_id = chunk.response_id if hasattr(chunk, "response_id") else None
if chunk.usage_metadata:
usage_data = self._extract_token_usage(chunk)
@@ -800,7 +806,7 @@ class GeminiCompletion(BaseLLM):
chunk=chunk.text,
from_task=from_task,
from_agent=from_agent,
- response_id=response_id
+ response_id=response_id,
)
if chunk.candidates:
@@ -837,7 +843,7 @@ class GeminiCompletion(BaseLLM):
"index": call_index,
},
call_type=LLMCallType.TOOL_CALL,
- response_id=response_id
+ response_id=response_id,
)
return full_response, function_calls, usage_data
@@ -972,7 +978,7 @@ class GeminiCompletion(BaseLLM):
from_task: Any | None = None,
from_agent: Any | None = None,
response_model: type[BaseModel] | None = None,
- ) -> str:
+ ) -> str | Any:
"""Handle streaming content generation."""
full_response = ""
function_calls: dict[int, dict[str, Any]] = {}
@@ -1050,7 +1056,7 @@ class GeminiCompletion(BaseLLM):
from_task: Any | None = None,
from_agent: Any | None = None,
response_model: type[BaseModel] | None = None,
- ) -> str:
+ ) -> str | Any:
"""Handle async streaming content generation."""
full_response = ""
function_calls: dict[int, dict[str, Any]] = {}
diff --git a/lib/crewai/tests/cassettes/llms/anthropic/test_anthropic_async_with_response_model.yaml b/lib/crewai/tests/cassettes/llms/anthropic/test_anthropic_async_with_response_model.yaml
index 34af261c2..03e805b8a 100644
--- a/lib/crewai/tests/cassettes/llms/anthropic/test_anthropic_async_with_response_model.yaml
+++ b/lib/crewai/tests/cassettes/llms/anthropic/test_anthropic_async_with_response_model.yaml
@@ -1,6 +1,8 @@
interactions:
- request:
- body: '{"max_tokens":4096,"messages":[{"role":"user","content":"Say hello in French"}],"model":"claude-sonnet-4-0","stream":false,"tool_choice":{"type":"tool","name":"structured_output"},"tools":[{"name":"structured_output","description":"Returns structured data according to the schema","input_schema":{"description":"Response model for greeting test.","properties":{"greeting":{"title":"Greeting","type":"string"},"language":{"title":"Language","type":"string"}},"required":["greeting","language"],"title":"GreetingResponse","type":"object"}}]}'
+ body: '{"max_tokens":4096,"messages":[{"role":"user","content":"Say hello in French"}],"model":"claude-sonnet-4-0","stream":false,"tool_choice":{"type":"tool","name":"structured_output"},"tools":[{"name":"structured_output","description":"Output
+ the structured response","input_schema":{"type":"object","description":"Response
+ model for greeting test.","title":"GreetingResponse","properties":{"greeting":{"type":"string","title":"Greeting"},"language":{"type":"string","title":"Language"}},"additionalProperties":false,"required":["greeting","language"]}}]}'
headers:
User-Agent:
- X-USER-AGENT-XXX
@@ -13,7 +15,7 @@ interactions:
connection:
- keep-alive
content-length:
- - '539'
+ - '551'
content-type:
- application/json
host:
@@ -29,7 +31,7 @@ interactions:
x-stainless-os:
- X-STAINLESS-OS-XXX
x-stainless-package-version:
- - 0.75.0
+ - 0.76.0
x-stainless-retry-count:
- '0'
x-stainless-runtime:
@@ -42,7 +44,7 @@ interactions:
uri: https://api.anthropic.com/v1/messages
response:
body:
- string: '{"model":"claude-sonnet-4-20250514","id":"msg_01XjvX2nCho1knuucbwwgCpw","type":"message","role":"assistant","content":[{"type":"tool_use","id":"toolu_019rfPRSDmBb7CyCTdGMv5rK","name":"structured_output","input":{"greeting":"Bonjour","language":"French"}}],"stop_reason":"tool_use","stop_sequence":null,"usage":{"input_tokens":432,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":53,"service_tier":"standard"}}'
+ string: '{"model":"claude-sonnet-4-20250514","id":"msg_01CKTyVmak15L5oQ36mv4sL9","type":"message","role":"assistant","content":[{"type":"tool_use","id":"toolu_0174BYmn6xiSnUwVhFD8S7EW","name":"structured_output","input":{"greeting":"Bonjour","language":"French"}}],"stop_reason":"tool_use","stop_sequence":null,"usage":{"input_tokens":436,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":53,"service_tier":"standard"}}'
headers:
CF-RAY:
- CF-RAY-XXX
@@ -51,7 +53,7 @@ interactions:
Content-Type:
- application/json
Date:
- - Mon, 01 Dec 2025 11:19:38 GMT
+ - Mon, 26 Jan 2026 14:59:34 GMT
Server:
- cloudflare
Transfer-Encoding:
@@ -82,12 +84,10 @@ interactions:
- DYNAMIC
request-id:
- REQUEST-ID-XXX
- retry-after:
- - '24'
strict-transport-security:
- STS-XXX
x-envoy-upstream-service-time:
- - '2101'
+ - '968'
status:
code: 200
message: OK
diff --git a/uv.lock b/uv.lock
index 36c4c60ba..499d4bd2d 100644
--- a/uv.lock
+++ b/uv.lock
@@ -310,7 +310,7 @@ wheels = [
[[package]]
name = "anthropic"
-version = "0.71.1"
+version = "0.73.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
{ name = "anyio" },
@@ -322,9 +322,9 @@ dependencies = [
{ name = "sniffio" },
{ name = "typing-extensions" },
]
-sdist = { url = "https://files.pythonhosted.org/packages/05/4b/19620875841f692fdc35eb58bf0201c8ad8c47b8443fecbf1b225312175b/anthropic-0.71.1.tar.gz", hash = "sha256:a77d156d3e7d318b84681b59823b2dee48a8ac508a3e54e49f0ab0d074e4b0da", size = 493294, upload-time = "2025-10-28T17:28:42.213Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/f0/07/f550112c3f5299d02f06580577f602e8a112b1988ad7c98ac1a8f7292d7e/anthropic-0.73.0.tar.gz", hash = "sha256:30f0d7d86390165f86af6ca7c3041f8720bb2e1b0e12a44525c8edfdbd2c5239", size = 425168, upload-time = "2025-11-14T18:47:52.635Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/4b/68/b2f988b13325f9ac9921b1e87f0b7994468014e1b5bd3bdbd2472f5baf45/anthropic-0.71.1-py3-none-any.whl", hash = "sha256:6ca6c579f0899a445faeeed9c0eb97aa4bdb751196262f9ccc96edfc0bb12679", size = 355020, upload-time = "2025-10-28T17:28:40.653Z" },
+ { url = "https://files.pythonhosted.org/packages/15/b1/5d4d3f649e151e58dc938cf19c4d0cd19fca9a986879f30fea08a7b17138/anthropic-0.73.0-py3-none-any.whl", hash = "sha256:0d56cd8b3ca3fea9c9b5162868bdfd053fbc189b8b56d4290bd2d427b56db769", size = 367839, upload-time = "2025-11-14T18:47:51.195Z" },
]
[[package]]
@@ -1276,7 +1276,7 @@ requires-dist = [
{ name = "aiobotocore", marker = "extra == 'aws'", specifier = "~=2.25.2" },
{ name = "aiocache", extras = ["memcached", "redis"], marker = "extra == 'a2a'", specifier = "~=0.12.3" },
{ name = "aiosqlite", specifier = "~=0.21.0" },
- { name = "anthropic", marker = "extra == 'anthropic'", specifier = "~=0.71.0" },
+ { name = "anthropic", marker = "extra == 'anthropic'", specifier = "~=0.73.0" },
{ name = "appdirs", specifier = "~=1.4.4" },
{ name = "azure-ai-inference", marker = "extra == 'azure-ai-inference'", specifier = "~=1.0.0b9" },
{ name = "boto3", marker = "extra == 'aws'", specifier = "~=1.40.38" },