planar 0.5.0__py3-none-any.whl → 0.8.0__py3-none-any.whl

planar/_version.py

@@ -1 +1 @@
-VERSION = "0.5.0"
+VERSION = "0.8.0"

planar/ai/agent.py

@@ -1,23 +1,14 @@
-from __future__ import annotations
-
 import inspect
-from dataclasses import dataclass, field
-from typing import (
-    Any,
-    Callable,
-    Dict,
-    List,
-    Type,
-    Union,
-    cast,
-    overload,
-)
+from dataclasses import dataclass
+from typing import Any, Type, cast
 
 from pydantic import BaseModel
+from pydantic_ai import models
 
+from planar.ai.agent_base import AgentBase
 from planar.ai.agent_utils import (
-    AgentEventEmitter,
     AgentEventType,
+    ModelSpec,
     ToolCallResult,
     create_tool_definition,
     extract_files_from_model,
@@ -25,171 +16,33 @@ from planar.ai.agent_utils import (
     render_template,
 )
 from planar.ai.models import (
-    AgentConfig,
     AgentRunResult,
     AssistantMessage,
-    CompletionResponse,
     ModelMessage,
     SystemMessage,
+    ToolDefinition,
+    ToolMessage,
     ToolResponse,
     UserMessage,
 )
-from planar.ai.providers import Anthropic, Gemini, Model, OpenAI
+from planar.ai.pydantic_ai import ModelRunResponse, model_run
 from planar.logging import get_logger
-from planar.modeling.field_helpers import JsonSchema
 from planar.utils import utc_now
-from planar.workflows import as_step
 from planar.workflows.models import StepType
 
 logger = get_logger(__name__)
 
 
-def _parse_model_string(model_str: str) -> Model:
-    """Parse a model string (e.g., 'openai:gpt-4.1') into a Model instance."""
-    parts = model_str.split(":", 1)
-    if len(parts) != 2:
-        raise ValueError(
-            f"Invalid model format: {model_str}. Expected format: 'provider:model_id'"
-        )
-
-    provider_id, model_id = parts
-
-    if provider_id.lower() == "openai":
-        return OpenAI.model(model_id)
-    elif provider_id.lower() == "anthropic":
-        return Anthropic.model(model_id)
-    elif provider_id.lower() == "gemini":
-        return Gemini.model(model_id)
-    else:
-        raise ValueError(f"Unsupported provider: {provider_id}")
-
-
 @dataclass
 class Agent[
-    # TODO: add `= str` default when we upgrade to 3.13
     TInput: BaseModel | str,
     TOutput: BaseModel | str,
-]:
-    """An LLM-powered agent that can be called directly within workflows."""
-
-    name: str
-    system_prompt: str
-    output_type: Type[TOutput] | None = None
-    input_type: Type[TInput] | None = None
-    user_prompt: str = ""
-    model: Union[str, Model] = "openai:gpt-4.1"
-    tools: List[Callable] = field(default_factory=list)
-    max_turns: int = 2
-    model_parameters: Dict[str, Any] = field(default_factory=dict)
-
-    # TODO: move here to serialize to frontend
-    #
-    # built_in_vars: Dict[str, str] = field(default_factory=lambda: {
-    #     "datetime_now": datetime.datetime.now().isoformat(),
-    #     "date_today": datetime.date.today().isoformat(),
-    # })
-
-    def __post_init__(self):
-        if self.input_type:
-            if (
-                not issubclass(self.input_type, BaseModel)
-                and self.input_type is not str
-            ):
-                raise ValueError(
-                    "input_type must be 'str' or a subclass of a Pydantic model"
-                )
-        if self.max_turns < 1:
-            raise ValueError("Max_turns must be greater than or equal to 1.")
-        if self.tools and self.max_turns <= 1:
-            raise ValueError(
-                "For tool calling to work, max_turns must be greater than 1."
-            )
-
-    def input_schema(self) -> JsonSchema | None:
-        if self.input_type is None:
-            return None
-        if self.input_type is str:
-            return None
-        assert issubclass(self.input_type, BaseModel), (
-            "input_type must be a subclass of BaseModel or str"
-        )
-        return self.input_type.model_json_schema()
-
-    def output_schema(self) -> JsonSchema | None:
-        if self.output_type is None:
-            return None
-        if self.output_type is str:
-            return None
-        assert issubclass(self.output_type, BaseModel), (
-            "output_type must be a subclass of BaseModel or str"
-        )
-        return self.output_type.model_json_schema()
-
-    def to_config(self) -> AgentConfig:
-        return AgentConfig(
-            system_prompt=self.system_prompt,
-            user_prompt=self.user_prompt,
-            model=str(self.model),
-            max_turns=self.max_turns,
-            model_parameters=self.model_parameters,
-        )
-
-    @overload
-    async def __call__(
-        self: "Agent[TInput, str]",
-        input_value: TInput,
-        event_emitter: AgentEventEmitter | None = None,
-    ) -> AgentRunResult[str]: ...
-
-    @overload
-    async def __call__(
-        self: "Agent[TInput, TOutput]",
-        input_value: TInput,
-        event_emitter: AgentEventEmitter | None = None,
-    ) -> AgentRunResult[TOutput]: ...
-
-    async def __call__(
-        self,
-        input_value: TInput,
-        event_emitter: AgentEventEmitter | None = None,
-    ) -> AgentRunResult[Any]:
-        if self.input_type is not None and not isinstance(input_value, self.input_type):
-            raise ValueError(
-                f"Input value must be of type {self.input_type}, but got {type(input_value)}"
-            )
-        elif not isinstance(input_value, (str, BaseModel)):
-            # Should not happen based on type constraints, but just in case
-            # user does not have type checking enabled
-            raise ValueError(
-                "Input value must be a string or a Pydantic model if input_type is not provided"
-            )
-
-        if self.output_type is None:
-            run_step = as_step(
-                self.run_step,
-                step_type=StepType.AGENT,
-                display_name=self.name,
-                return_type=AgentRunResult[str],
-            )
-        else:
-            run_step = as_step(
-                self.run_step,
-                step_type=StepType.AGENT,
-                display_name=self.name,
-                return_type=AgentRunResult[self.output_type],
-            )
-
-        result = await run_step(
-            input_value=input_value,
-            event_emitter=event_emitter,
-        )
-        # Cast the result to ensure type compatibility
-        return cast(AgentRunResult[TOutput], result)
+](AgentBase[TInput, TOutput]):
+    model: models.KnownModelName | models.Model = "openai:gpt-4o"
 
     async def run_step(
         self,
         input_value: TInput,
-        event_emitter: AgentEventEmitter | None = None,
     ) -> AgentRunResult[TOutput]:
         """Execute the agent with the provided inputs.
 
@@ -200,6 +53,7 @@ class Agent[
         Returns:
             AgentRunResult containing the agent's response
         """
+        event_emitter = self.event_emitter
         logger.debug(
             "agent run_step called", agent_name=self.name, input_type=type(input_value)
         )
@@ -264,18 +118,18 @@ class Agent[
             raise ValueError(f"Missing required parameter for prompt formatting: {e}")
 
         # Get the LLM provider and model
-        model_config = config.model
-        if isinstance(model_config, str):
-            model = _parse_model_string(model_config)
+        if isinstance(self.model, str):
+            model = models.infer_model(self.model)
         else:
-            model = model_config
+            model = self.model
 
         # Apply model parameters if specified
+        model_settings = None
         if config.model_parameters:
-            model = model.with_parameters(**config.model_parameters)
+            model_settings = config.model_parameters
 
         # Prepare structured messages
-        messages: List[ModelMessage] = []
+        messages: list[ModelMessage] = []
         if formatted_system_prompt:
             messages.append(SystemMessage(content=formatted_system_prompt))
 
@@ -287,167 +141,182 @@ class Agent[
         if self.tools:
             tool_definitions = [create_tool_definition(tool) for tool in self.tools]
 
-        # Determine output type for the provider call
+        # Determine output type for the agent call
         # Pass the Pydantic model type if output_type is a subclass of BaseModel,
         # otherwise pass None (indicating string output is expected).
-        output_type_for_provider: Type[BaseModel] | None = None
+        output_type: Type[BaseModel] | None = None
         # Use issubclass safely by checking if output_type is a type first
         if inspect.isclass(self.output_type) and issubclass(
             self.output_type, BaseModel
         ):
-            output_type_for_provider = cast(Type[BaseModel], self.output_type)
+            output_type = cast(Type[BaseModel], self.output_type)
 
         # Execute the LLM call
         max_turns = config.max_turns
 
-        # Single turn completion (default case)
-        result = None
-        if not tool_definitions:
+        # We use this inner function to pass "model" and "event_emitter",
+        # which are not serializable as step parameters.
+        async def agent_run_step(
+            model_spec: ModelSpec,
+            messages: list[ModelMessage],
+            turns_left: int,
+            tools: list[ToolDefinition] | None = None,
+            output_type: Type[BaseModel] | None = None,
+        ):
             logger.debug(
-                "agent performing single turn completion",
+                "agent running",
                 agent_name=self.name,
-                model=model.model_spec,
-                output_type=output_type_for_provider,
+                model=model_spec,
+                model_settings=model_settings,
+                output_type=output_type,
             )
-            response = await as_step(
-                model.provider_class.complete,
+            if output_type is None:
+                return await model_run(
+                    model=model,
+                    max_extra_turns=turns_left,
+                    model_settings=model_settings,
+                    messages=messages,
+                    tools=tools or [],
+                    event_handler=cast(Any, event_emitter),
+                )
+            else:
+                return await model_run(
+                    model=model,
+                    max_extra_turns=turns_left,
+                    model_settings=model_settings,
+                    messages=messages,
+                    output_type=output_type,
+                    tools=tools or [],
+                    event_handler=cast(Any, event_emitter),
+                )
+
+        model_spec = ModelSpec(
+            model_id=str(model),
+            parameters=config.model_parameters,
+        )
+        result = None
+        logger.debug(
+            "agent performing multi-turn completion with tools",
+            agent_name=self.name,
+            max_turns=max_turns,
+        )
+        turns_left = max_turns
+        while turns_left > 0:
+            turns_left -= 1
+            logger.debug("agent turn", agent_name=self.name, turns_left=turns_left)
+
+            # Get model response
+            run_response = await self.as_step_if_durable(
+                agent_run_step,
                 step_type=StepType.AGENT,
-                return_type=CompletionResponse[output_type_for_provider or str],
+                return_type=ModelRunResponse[output_type or str],
             )(
-                model_spec=model.model_spec,
+                model_spec=model_spec,
                 messages=messages,
-                output_type=output_type_for_provider,
+                turns_left=turns_left,
+                output_type=output_type,
+                tools=tool_definitions or [],
             )
-            result = response.content
+            response = run_response.response
+            turns_left -= run_response.extra_turns_used
 
             # Emit response event if event_emitter is provided
             if event_emitter:
                 event_emitter.emit(AgentEventType.RESPONSE, response.content)
-        else:
+
+            # If no tool calls or last turn, return content
+            if not response.tool_calls or turns_left == 0:
+                logger.debug(
+                    "agent completion: no tool calls or last turn",
+                    agent_name=self.name,
+                    has_content=response.content is not None,
+                )
+                result = response.content
+                break
+
+            # Process tool calls
             logger.debug(
-                "agent performing multi-turn completion with tools",
+                "agent received tool calls",
                 agent_name=self.name,
-                max_turns=max_turns,
+                num_tool_calls=len(response.tool_calls),
             )
-            # Multi-turn with tools
-            turns_left = max_turns
-            while turns_left > 0:
-                turns_left -= 1
-                logger.debug("agent turn", agent_name=self.name, turns_left=turns_left)
-
-                # Get model response
-                response = await as_step(
-                    model.provider_class.complete,
-                    step_type=StepType.AGENT,
-                    return_type=CompletionResponse[output_type_for_provider or str],
-                )(
-                    model_spec=model.model_spec,
-                    messages=messages,
-                    output_type=output_type_for_provider,
-                    tools=tool_definitions,
-                )
-
-                # Emit response event if event_emitter is provided
-                if event_emitter:
-                    event_emitter.emit(AgentEventType.RESPONSE, response.content)
-
-                # If no tool calls or last turn, return content
-                if not response.tool_calls or turns_left == 0:
-                    logger.debug(
-                        "agent completion: no tool calls or last turn",
-                        agent_name=self.name,
-                        has_content=response.content is not None,
-                    )
-                    result = response.content
-                    break
+            assistant_message = AssistantMessage(
+                content=None,
+                tool_calls=response.tool_calls,
+            )
+            messages.append(assistant_message)
 
-                # Process tool calls
+            # Execute each tool and add tool responses to messages
+            for tool_call_idx, tool_call in enumerate(response.tool_calls):
                 logger.debug(
-                    "agent received tool calls",
+                    "agent processing tool call",
                     agent_name=self.name,
-                    num_tool_calls=len(response.tool_calls),
+                    tool_call_index=tool_call_idx + 1,
+                    tool_call_id=tool_call.id,
+                    tool_call_name=tool_call.name,
                 )
-                assistant_message = AssistantMessage(
-                    content=None,
-                    tool_calls=response.tool_calls,
+                # Find the matching tool function
+                tool_fn = next(
+                    (t for t in self.tools if t.__name__ == tool_call.name),
+                    None,
                 )
-                messages.append(assistant_message)
 
-                # Execute each tool and add tool responses to messages
-                for tool_call_idx, tool_call in enumerate(response.tool_calls):
-                    logger.debug(
-                        "agent processing tool call",
+                if not tool_fn:
+                    tool_result = f"Error: Tool '{tool_call.name}' not found."
+                    logger.warning(
+                        "tool not found for agent",
+                        tool_name=tool_call.name,
                         agent_name=self.name,
-                        tool_call_index=tool_call_idx + 1,
-                        tool_call_id=tool_call.id,
-                        tool_call_name=tool_call.name,
                     )
-                    # Find the matching tool function
-                    tool_fn = next(
-                        (t for t in self.tools if t.__name__ == tool_call.name),
-                        None,
+                else:
+                    # Execute the tool with the provided arguments
+                    tool_result = await self.as_step_if_durable(
+                        tool_fn,
+                        step_type=StepType.TOOL_CALL,
+                    )(**tool_call.arguments)
+                    logger.info(
+                        "tool executed by agent",
+                        tool_name=tool_call.name,
+                        agent_name=self.name,
+                        result_type=type(tool_result),
                     )
 
-                    if not tool_fn:
-                        tool_result = f"Error: Tool '{tool_call.name}' not found."
-                        logger.warning(
-                            "tool not found for agent",
-                            tool_name=tool_call.name,
-                            agent_name=self.name,
-                        )
-                    else:
-                        # Execute the tool with the provided arguments
-                        tool_result = await as_step(
-                            tool_fn,
-                            step_type=StepType.TOOL_CALL,
-                        )(**tool_call.arguments)
-                        logger.info(
-                            "tool executed by agent",
-                            tool_name=tool_call.name,
-                            agent_name=self.name,
-                            result_type=type(tool_result),
-                        )
-
-                    # Create a tool response
-                    tool_response = ToolResponse(
-                        tool_call_id=tool_call.id or "call_1", content=str(tool_result)
-                    )
+                # Create a tool response
+                tool_response = ToolResponse(
+                    tool_call_id=tool_call.id or "call_1", content=str(tool_result)
+                )
 
-                    # Emit tool response event if event_emitter is provided
-                    if event_emitter:
-                        event_emitter.emit(
-                            AgentEventType.TOOL_RESPONSE,
-                            ToolCallResult(
-                                tool_call_id=tool_call.id or "call_1",
-                                tool_call_name=tool_call.name,
-                                content=tool_result,
-                            ),
-                        )
-
-                    # Convert the tool response to a message based on provider
-                    tool_message = model.provider_class.format_tool_response(
-                        tool_response
+                # Emit tool response event if event_emitter is provided
+                if event_emitter:
+                    event_emitter.emit(
+                        AgentEventType.TOOL_RESPONSE,
+                        ToolCallResult(
+                            tool_call_id=tool_call.id or "call_1",
+                            tool_call_name=tool_call.name,
+                            content=tool_result,
+                        ),
                     )
-                    messages.append(tool_message)
 
-                # Continue to next turn
-
-            if result is None:
-                logger.warning(
-                    "agent completed tool interactions but result is none",
-                    agent_name=self.name,
-                    expected_type=self.output_type,
-                )
-                raise ValueError(
-                    f"Expected result of type {self.output_type} but got none after tool interactions."
+                tool_message = ToolMessage(
+                    content=tool_response.content,
+                    tool_call_id=tool_response.tool_call_id or "call_1",
                 )
+                messages.append(tool_message)
 
-        if event_emitter:
-            event_emitter.emit(AgentEventType.COMPLETED, result)
+            # Continue to next turn
 
         if result is None:
-            logger.warning("agent final result is none", agent_name=self.name)
-            raise ValueError("No result obtained after tool interactions")
+            logger.warning(
+                "agent completed tool interactions but result is none",
+                agent_name=self.name,
+                expected_type=self.output_type,
+            )
+            raise ValueError(
+                f"Expected result of type {self.output_type} but got none after tool interactions."
+            )
+
+        if event_emitter:
+            event_emitter.emit(AgentEventType.COMPLETED, result)
 
         logger.info(
             "agent completed",
@@ -455,3 +324,6 @@ class Agent[
             final_result_type=type(result),
         )
         return AgentRunResult[TOutput](output=cast(TOutput, result))
+
+    def get_model_str(self) -> str:
+        return str(self.model)