18530018322

Committed 15 Oct 2025 01:10PM UTC coverage: 92.075% (-0.03%) from 92.103%

Build # 18530018322

Build Type

Pull #9880

github

Committed by

web-flow

Commit Message

Merge 6dad544fe into cfa5d2761

Pull Request Pull Request #9880: draft: Expand tools param to include list[Toolset]

Run Details

13279 of 14422 relevant lines covered (92.07%)

0.92 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

95.9

haystack/components/tools/tool_invoker.py

# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
#
# SPDX-License-Identifier: Apache-2.0

import asyncio
import contextvars
import inspect
import json
from concurrent.futures import ThreadPoolExecutor
from functools import partial
from typing import Any, Callable, Optional, Union

from haystack.components.agents import State
from haystack.core.component.component import component
from haystack.core.component.sockets import Sockets
from haystack.core.serialization import default_from_dict, default_to_dict, logging
from haystack.dataclasses import ChatMessage, ToolCall
from haystack.dataclasses.streaming_chunk import StreamingCallbackT, StreamingChunk, select_streaming_callback
from haystack.tools import (
    ComponentTool,
    Tool,
    Toolset,
    _check_duplicate_tool_names,
    deserialize_tools_or_toolset_inplace,
    flatten_tools_or_toolsets,
    serialize_tools_or_toolset,
)
from haystack.tools.errors import ToolInvocationError
from haystack.tracing.utils import _serializable_value
from haystack.utils.callable_serialization import deserialize_callable, serialize_callable

logger = logging.getLogger(__name__)


class ToolInvokerError(Exception):
    """Base exception class for ToolInvoker errors."""

    def __init__(self, message: str):
        super().__init__(message)


class ToolNotFoundException(ToolInvokerError):
    """Exception raised when a tool is not found in the list of available tools."""

    def __init__(self, tool_name: str, available_tools: list[str]):
        message = f"Tool '{tool_name}' not found. Available tools: {', '.join(available_tools)}"
        super().__init__(message)


class StringConversionError(ToolInvokerError):
    """Exception raised when the conversion of a tool result to a string fails."""

    def __init__(self, tool_name: str, conversion_function: str, error: Exception):
        message = f"Failed to convert tool result from tool {tool_name} using '{conversion_function}'. Error: {error}"
        super().__init__(message)


class ToolOutputMergeError(ToolInvokerError):
    """Exception raised when merging tool outputs into state fails."""

    @classmethod
    def from_exception(cls, tool_name: str, error: Exception) -> "ToolOutputMergeError":
        """
        Create a ToolOutputMergeError from an exception.
        """
        message = f"Failed to merge tool outputs from tool {tool_name} into State: {error}"
        return cls(message)


@component
class ToolInvoker:
    """
    Invokes tools based on prepared tool calls and returns the results as a list of ChatMessage objects.

    Also handles reading/writing from a shared `State`.
    At initialization, the ToolInvoker component is provided with a list of available tools.
    At runtime, the component processes a list of ChatMessage object containing tool calls
    and invokes the corresponding tools.
    The results of the tool invocations are returned as a list of ChatMessage objects with tool role.

    Usage example:
    ```python
    from haystack.dataclasses import ChatMessage, ToolCall
    from haystack.tools import Tool
    from haystack.components.tools import ToolInvoker

    # Tool definition
    def dummy_weather_function(city: str):
        return f"The weather in {city} is 20 degrees."

    parameters = {"type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"]}

    tool = Tool(name="weather_tool",
                description="A tool to get the weather",
                function=dummy_weather_function,
                parameters=parameters)

    # Usually, the ChatMessage with tool_calls is generated by a Language Model
    # Here, we create it manually for demonstration purposes
    tool_call = ToolCall(
        tool_name="weather_tool",
        arguments={"city": "Berlin"}
    )
    message = ChatMessage.from_assistant(tool_calls=[tool_call])

    # ToolInvoker initialization and run
    invoker = ToolInvoker(tools=[tool])
    result = invoker.run(messages=[message])

    print(result)
    ```

    ```
    >>  {
    >>      'tool_messages': [
    >>          ChatMessage(
    >>              _role=<ChatRole.TOOL: 'tool'>,
    >>              _content=[
    >>                  ToolCallResult(
    >>                      result='"The weather in Berlin is 20 degrees."',
    >>                      origin=ToolCall(
    >>                          tool_name='weather_tool',
    >>                          arguments={'city': 'Berlin'},
    >>                          id=None
    >>                      )
    >>                  )
    >>              ],
    >>              _meta={}
    >>          )
    >>      ]
    >>  }
    ```

    Usage example with a Toolset:
    ```python
    from haystack.dataclasses import ChatMessage, ToolCall
    from haystack.tools import Tool, Toolset
    from haystack.components.tools import ToolInvoker

    # Tool definition
    def dummy_weather_function(city: str):
        return f"The weather in {city} is 20 degrees."

    parameters = {"type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"]}

    tool = Tool(name="weather_tool",
                description="A tool to get the weather",
                function=dummy_weather_function,
                parameters=parameters)

    # Create a Toolset
    toolset = Toolset([tool])

    # Usually, the ChatMessage with tool_calls is generated by a Language Model
    # Here, we create it manually for demonstration purposes
    tool_call = ToolCall(
        tool_name="weather_tool",
        arguments={"city": "Berlin"}
    )
    message = ChatMessage.from_assistant(tool_calls=[tool_call])

    # ToolInvoker initialization and run with Toolset
    invoker = ToolInvoker(tools=toolset)
    result = invoker.run(messages=[message])

    print(result)
    """

    def __init__(
        self,
        tools: Union[list[Tool], Toolset, list[Toolset]],
        raise_on_failure: bool = True,
        convert_result_to_json_string: bool = False,
        streaming_callback: Optional[StreamingCallbackT] = None,
        *,
        enable_streaming_callback_passthrough: bool = False,
        max_workers: int = 4,
    ):
        """
        Initialize the ToolInvoker component.

        :param tools:
            A list of tools, a Toolset instance, or a list of Toolset instances that can resolve tools.
        :param raise_on_failure:
            If True, the component will raise an exception in case of errors
            (tool not found, tool invocation errors, tool result conversion errors).
            If False, the component will return a ChatMessage object with `error=True`
            and a description of the error in `result`.
        :param convert_result_to_json_string:
            If True, the tool invocation result will be converted to a string using `json.dumps`.
            If False, the tool invocation result will be converted to a string using `str`.
        :param streaming_callback:
            A callback function that will be called to emit tool results.
            Note that the result is only emitted once it becomes available — it is not
            streamed incrementally in real time.
        :param enable_streaming_callback_passthrough:
            If True, the `streaming_callback` will be passed to the tool invocation if the tool supports it.
            This allows tools to stream their results back to the client.
            Note that this requires the tool to have a `streaming_callback` parameter in its `invoke` method signature.
            If False, the `streaming_callback` will not be passed to the tool invocation.
        :param max_workers:
            The maximum number of workers to use in the thread pool executor.
            This also decides the maximum number of concurrent tool invocations.
        :raises ValueError:
            If no tools are provided or if duplicate tool names are found.
        """
        self.tools = tools
        self.streaming_callback = streaming_callback
        self.enable_streaming_callback_passthrough = enable_streaming_callback_passthrough
        self.max_workers = max_workers
        self.raise_on_failure = raise_on_failure
        self.convert_result_to_json_string = convert_result_to_json_string

        self._tools_with_names = self._validate_and_prepare_tools(tools)

    @staticmethod
    def _make_context_bound_invoke(tool_to_invoke: Tool, final_args: dict[str, Any]) -> Callable[[], Any]:
        """
        Create a zero-arg callable that invokes the tool under the caller's contextvars Context.

        We copy and use contextvars to preserve the caller’s ambient execution context (for example the active
        tracing Span) across thread boundaries. Python’s contextvars do not automatically propagate to worker
        threads (or to threadpool tasks spawned via run_in_executor), so without intervention nested tool calls
        would lose their parent trace/span and appear as separate roots. By capturing the current Context in the
        caller thread and invoking the tool under ctx.run(...) inside the executor, we ensure proper span parentage,
        consistent tagging, and reliable log/trace correlation in both sync and async paths. The callable returns
        ToolInvocationError instead of raising so parallel execution can collect failures.
        """
        ctx = contextvars.copy_context()

        def _runner() -> Any:
            try:
                return ctx.run(partial(tool_to_invoke.invoke, **final_args))
            except ToolInvocationError as e:
                return e

        return _runner

    @staticmethod
    def _validate_and_prepare_tools(tools: Union[list[Tool], Toolset, list[Toolset]]) -> dict[str, Tool]:
        """
        Validates and prepares tools for use by the ToolInvoker.

        :param tools: A list of tools or a Toolset instance.
        :returns: A dictionary mapping tool names to Tool instances.
        :raises ValueError: If no tools are provided or if duplicate tool names are found.
        """
        if not tools:
            raise ValueError("ToolInvoker requires at least one tool.")

        converted_tools = flatten_tools_or_toolsets(tools)

        _check_duplicate_tool_names(converted_tools)
        tool_names = [tool.name for tool in converted_tools]
        duplicates = {name for name in tool_names if tool_names.count(name) > 1}
        if duplicates:
            raise ValueError(f"Duplicate tool names found: {duplicates}")

        return dict(zip(tool_names, converted_tools))

    def _default_output_to_string_handler(self, result: Any) -> str:
        """
        Default handler for converting a tool result to a string.

        :param result: The tool result to convert to a string.
        :returns: The converted tool result as a string.
        """
        # We iterate through all items in result and call to_dict() if present
        # Relevant for a few reasons:
        # - If using convert_result_to_json_string we'd rather convert Haystack objects to JSON serializable dicts
        # - If using default str() we prefer converting Haystack objects to dicts rather than relying on the
        #   __repr__ method
        serializable = _serializable_value(result)

        if self.convert_result_to_json_string:
            try:
                # We disable ensure_ascii so special chars like emojis are not converted
                str_result = json.dumps(serializable, ensure_ascii=False)
            except Exception as error:
                # If the result is not JSON serializable, we fall back to str
                logger.warning(
                    "Tool result is not JSON serializable. Falling back to str conversion. "
                    "Result: {result}\nError: {error}",
                    result=result,
                    err=error,
                )
                str_result = str(result)
            return str_result

        return str(serializable)

    def _prepare_tool_result_message(self, result: Any, tool_call: ToolCall, tool_to_invoke: Tool) -> ChatMessage:
        """
        Prepares a ChatMessage with the result of a tool invocation.

        :param result:
            The tool result.
        :param tool_call:
            The ToolCall object containing the tool name and arguments.
        :param tool_to_invoke:
            The Tool object that was invoked.
        :returns:
            A ChatMessage object containing the tool result as a string.
        :raises
            StringConversionError: If the conversion of the tool result to a string fails
            and `raise_on_failure` is True.
        """
        outputs_config = tool_to_invoke.outputs_to_string or {}
        source_key = outputs_config.get("source")

        # If no handler is provided, we use the default handler
        output_to_string_handler = outputs_config.get("handler", self._default_output_to_string_handler)

        # If a source key is provided, we extract the result from the source key
        result_to_convert = result.get(source_key) if source_key is not None else result

        try:
            tool_result_str = output_to_string_handler(result_to_convert)
            chat_message = ChatMessage.from_tool(tool_result=tool_result_str, origin=tool_call)
        except Exception as e:
            error = StringConversionError(tool_call.tool_name, output_to_string_handler.__name__, e)
            if self.raise_on_failure:
                raise error from e
            logger.error("{error_exception}", error_exception=error)
            chat_message = ChatMessage.from_tool(tool_result=str(error), origin=tool_call, error=True)
        return chat_message

    @staticmethod
    def _get_func_params(tool: Tool) -> set:
        """
        Returns the function parameters of the tool's invoke method.

        This method inspects the tool's function signature to determine which parameters the tool accepts.
        """
        # ComponentTool wraps the function with a function that accepts kwargs, so we need to look at input sockets
        # to find out which parameters the tool accepts.
        if isinstance(tool, ComponentTool):
            # mypy doesn't know that ComponentMeta always adds __haystack_input__ to Component
            assert hasattr(tool._component, "__haystack_input__") and isinstance(
                tool._component.__haystack_input__, Sockets
            )
            func_params = set(tool._component.__haystack_input__._sockets_dict.keys())
        else:
            func_params = set(inspect.signature(tool.function).parameters.keys())

        return func_params

    @staticmethod
    def _inject_state_args(tool: Tool, llm_args: dict[str, Any], state: State) -> dict[str, Any]:
        """
        Combine LLM-provided arguments (llm_args) with state-based arguments.

        Tool arguments take precedence in the following order:
          - LLM overrides state if the same param is present in both
          - local tool.inputs_from_state mappings (if any)
          - function signature name matching
        """
        final_args = dict(llm_args)  # start with LLM-provided
        func_params = ToolInvoker._get_func_params(tool)

        # Determine the source of parameter mappings (explicit tool inputs or direct function parameters)
        # Typically, a "Tool" might have .inputs_from_state = {"state_key": "tool_param_name"}
        if hasattr(tool, "inputs_from_state") and isinstance(tool.inputs_from_state, dict):
            param_mappings = tool.inputs_from_state
        else:
            param_mappings = {name: name for name in func_params}

        # Populate final_args from state if not provided by LLM
        for state_key, param_name in param_mappings.items():
            if param_name not in final_args and state.has(state_key):
                final_args[param_name] = state.get(state_key)

        return final_args

    @staticmethod
    def _merge_tool_outputs(tool: Tool, result: Any, state: State) -> None:
        """
        Merges the tool result into the State.

        This method processes the output of a tool execution and integrates it into the global state.
        It also determines what message, if any, should be returned for further processing in a conversation.

        Processing Steps:
        1. If `result` is not a dictionary, nothing is stored into state and the full `result` is returned.
        2. If the `tool` does not define an `outputs_to_state` mapping nothing is stored into state.
           The return value in this case is simply the full `result` dictionary.
        3. If the tool defines an `outputs_to_state` mapping (a dictionary describing how the tool's output should be
           processed), the method delegates to `_handle_tool_outputs` to process the output accordingly.
           This allows certain fields in `result` to be mapped explicitly to state fields or formatted using custom
           handlers.

        :param tool: Tool instance containing optional `outputs_to_state` mapping to guide result processing.
        :param result: The output from tool execution. Can be a dictionary, or any other type.
        :param state: The global State object to which results should be merged.
        :returns: Three possible values:
            - A string message for conversation
            - The merged result dictionary
            - Or the raw result if not a dictionary
        """
        # If result is not a dictionary we exit
        if not isinstance(result, dict):
            return

        # If there is no specific `outputs_to_state` mapping, we exit
        if not hasattr(tool, "outputs_to_state") or not isinstance(tool.outputs_to_state, dict):
            return

        # Update the state with the tool outputs
        for state_key, config in tool.outputs_to_state.items():
            # Get the source key from the output config, otherwise use the entire result
            source_key = config.get("source", None)
            output_value = result.get(source_key) if source_key else result

            # Merge other outputs into the state
            state.set(state_key, output_value, handler_override=config.get("handler"))

    @staticmethod
    def _create_tool_result_streaming_chunk(tool_messages: list[ChatMessage], tool_call: ToolCall) -> StreamingChunk:
        """Create a streaming chunk for a tool result."""
        return StreamingChunk(
            content="",
            index=len(tool_messages) - 1,
            tool_call_result=tool_messages[-1].tool_call_results[0],
            start=True,
            meta={"tool_result": tool_messages[-1].tool_call_results[0].result, "tool_call": tool_call},
        )

    def _prepare_tool_call_params(
        self,
        *,
        messages_with_tool_calls: list[ChatMessage],
        state: State,
        streaming_callback: Optional[StreamingCallbackT],
        enable_streaming_passthrough: bool,
        tools_with_names: dict[str, Tool],
    ) -> tuple[list[ToolCall], list[dict[str, Any]], list[ChatMessage]]:
        """
        Prepare tool call parameters for execution and collect any error messages.

        :param messages_with_tool_calls: Messages containing tool calls to process
        :param state: The current state for argument injection
        :param streaming_callback: Optional streaming callback to inject
        :param enable_streaming_passthrough: Whether to pass streaming callback to tools
        :returns: Tuple of (tool_calls, tool_call_params, error_messages)
        """
        tool_call_params = []
        error_messages = []
        tool_calls = []

        for message in messages_with_tool_calls:
            for tool_call in message.tool_calls:
                tool_name = tool_call.tool_name

                # Check if the tool is available, otherwise return an error message
                if tool_name not in tools_with_names:
                    error = ToolNotFoundException(tool_name, list(tools_with_names.keys()))
                    if self.raise_on_failure:
                        raise error
                    logger.error("{error_exception}", error_exception=error)
                    error_messages.append(ChatMessage.from_tool(tool_result=str(error), origin=tool_call, error=True))
                    continue

                tool_to_invoke = tools_with_names[tool_name]

                # Combine user + state inputs
                llm_args = tool_call.arguments.copy()
                final_args = self._inject_state_args(tool_to_invoke, llm_args, state)

                # Check whether to inject streaming_callback
                if (
                    enable_streaming_passthrough
                    and streaming_callback is not None
                    and "streaming_callback" not in final_args
                    and "streaming_callback" in self._get_func_params(tool_to_invoke)
                ):
                    final_args["streaming_callback"] = streaming_callback

                tool_call_params.append({"tool_to_invoke": tool_to_invoke, "final_args": final_args})
                tool_calls.append(tool_call)

        return tool_calls, tool_call_params, error_messages

    @component.output_types(tool_messages=list[ChatMessage], state=State)
    def run(
        self,
        messages: list[ChatMessage],
        state: Optional[State] = None,
        streaming_callback: Optional[StreamingCallbackT] = None,
        *,
        enable_streaming_callback_passthrough: Optional[bool] = None,
        tools: Optional[Union[list[Tool], Toolset, list[Toolset]]] = None,
    ) -> dict[str, Any]:
        """
        Processes ChatMessage objects containing tool calls and invokes the corresponding tools, if available.

        :param messages:
            A list of ChatMessage objects.
        :param state: The runtime state that should be used by the tools.
        :param streaming_callback: A callback function that will be called to emit tool results.
            Note that the result is only emitted once it becomes available — it is not
            streamed incrementally in real time.
        :param enable_streaming_callback_passthrough:
            If True, the `streaming_callback` will be passed to the tool invocation if the tool supports it.
            This allows tools to stream their results back to the client.
            Note that this requires the tool to have a `streaming_callback` parameter in its `invoke` method signature.
            If False, the `streaming_callback` will not be passed to the tool invocation.
            If None, the value from the constructor will be used.
        :param tools:
            A list of tools to use for the tool invoker. If set, overrides the tools set in the constructor.
        :returns:
            A dictionary with the key `tool_messages` containing a list of ChatMessage objects with tool role.
            Each ChatMessage objects wraps the result of a tool invocation.

        :raises ToolNotFoundException:
            If the tool is not found in the list of available tools and `raise_on_failure` is True.
        :raises ToolInvocationError:
            If the tool invocation fails and `raise_on_failure` is True.
        :raises StringConversionError:
            If the conversion of the tool result to a string fails and `raise_on_failure` is True.
        :raises ToolOutputMergeError:
            If merging tool outputs into state fails and `raise_on_failure` is True.
        """
        tools_with_names = self._tools_with_names
        if tools is not None:
            tools_with_names = self._validate_and_prepare_tools(tools)
            logger.debug(
                f"For this invocation, overriding constructor tools with: {', '.join(tools_with_names.keys())}"
            )

        if state is None:
            state = State(schema={})

        resolved_enable_streaming_passthrough = (
            enable_streaming_callback_passthrough
            if enable_streaming_callback_passthrough is not None
            else self.enable_streaming_callback_passthrough
        )

        # Only keep messages with tool calls
        messages_with_tool_calls = [message for message in messages if message.tool_calls]
        streaming_callback = select_streaming_callback(
            init_callback=self.streaming_callback, runtime_callback=streaming_callback, requires_async=False
        )

        if not messages_with_tool_calls:
            return {"tool_messages": [], "state": state}

        # 1) Collect all tool calls and their parameters for parallel execution
        tool_messages = []
        tool_calls, tool_call_params, error_messages = self._prepare_tool_call_params(
            messages_with_tool_calls=messages_with_tool_calls,
            state=state,
            streaming_callback=streaming_callback,
            enable_streaming_passthrough=resolved_enable_streaming_passthrough,
            tools_with_names=tools_with_names,
        )
        tool_messages.extend(error_messages)

        if not tool_call_params:
            return {"tool_messages": tool_messages, "state": state}

        # 2) Execute valid tool calls in parallel
        with ThreadPoolExecutor(max_workers=self.max_workers) as executor:
            futures = []
            for params in tool_call_params:
                callable_ = self._make_context_bound_invoke(params["tool_to_invoke"], params["final_args"])
                futures.append(executor.submit(callable_))

            # 3) Gather and process results: handle errors and merge outputs into state
            for future, tool_call in zip(futures, tool_calls):
                result = future.result()

                if isinstance(result, ToolInvocationError):
                    # a) This is an error, create error Tool message
                    if self.raise_on_failure:
                        raise result
                    logger.error("{error_exception}", error_exception=result)
                    tool_messages.append(ChatMessage.from_tool(tool_result=str(result), origin=tool_call, error=True))
                else:
                    # b) In case of success, merge outputs into state
                    try:
                        tool_to_invoke = tools_with_names[tool_call.tool_name]
                        self._merge_tool_outputs(tool=tool_to_invoke, result=result, state=state)
                        tool_messages.append(
                            self._prepare_tool_result_message(
                                result=result, tool_call=tool_call, tool_to_invoke=tool_to_invoke
                            )
                        )
                    except Exception as e:
                        error = ToolOutputMergeError.from_exception(tool_name=tool_call.tool_name, error=e)
                        if self.raise_on_failure:
                            raise error from e
                        logger.error("{error_exception}", error_exception=error)
                        tool_messages.append(
                            ChatMessage.from_tool(tool_result=str(error), origin=tool_call, error=True)
                        )

                # c) Handle streaming callback
                if streaming_callback is not None:
                    streaming_callback(
                        self._create_tool_result_streaming_chunk(tool_messages=tool_messages, tool_call=tool_call)
                    )

        # We stream one more chunk that contains a finish_reason if tool_messages were generated
        if len(tool_messages) > 0 and streaming_callback is not None:
            streaming_callback(
                StreamingChunk(
                    content="", finish_reason="tool_call_results", meta={"finish_reason": "tool_call_results"}
                )
            )

        return {"tool_messages": tool_messages, "state": state}

    @component.output_types(tool_messages=list[ChatMessage], state=State)
    async def run_async(
        self,
        messages: list[ChatMessage],
        state: Optional[State] = None,
        streaming_callback: Optional[StreamingCallbackT] = None,
        *,
        enable_streaming_callback_passthrough: Optional[bool] = None,
        tools: Optional[Union[list[Tool], Toolset, list[Toolset]]] = None,
    ) -> dict[str, Any]:
        """
        Asynchronously processes ChatMessage objects containing tool calls.

        Multiple tool calls are performed concurrently.
        :param messages:
            A list of ChatMessage objects.
        :param state: The runtime state that should be used by the tools.
        :param streaming_callback: An asynchronous callback function that will be called to emit tool results.
            Note that the result is only emitted once it becomes available — it is not
            streamed incrementally in real time.
        :param enable_streaming_callback_passthrough:
            If True, the `streaming_callback` will be passed to the tool invocation if the tool supports it.
            This allows tools to stream their results back to the client.
            Note that this requires the tool to have a `streaming_callback` parameter in its `invoke` method signature.
            If False, the `streaming_callback` will not be passed to the tool invocation.
            If None, the value from the constructor will be used.
        :param tools:
            A list of tools to use for the tool invoker. If set, overrides the tools set in the constructor.
        :returns:
            A dictionary with the key `tool_messages` containing a list of ChatMessage objects with tool role.
            Each ChatMessage objects wraps the result of a tool invocation.

        :raises ToolNotFoundException:
            If the tool is not found in the list of available tools and `raise_on_failure` is True.
        :raises ToolInvocationError:
            If the tool invocation fails and `raise_on_failure` is True.
        :raises StringConversionError:
            If the conversion of the tool result to a string fails and `raise_on_failure` is True.
        :raises ToolOutputMergeError:
            If merging tool outputs into state fails and `raise_on_failure` is True.
        """

        tools_with_names = self._tools_with_names
        if tools is not None:
            tools_with_names = self._validate_and_prepare_tools(tools)
            logger.debug(
                f"For this invocation, overriding constructor tools with: {', '.join(tools_with_names.keys())}"
            )

        if state is None:
            state = State(schema={})

        resolved_enable_streaming_passthrough = (
            enable_streaming_callback_passthrough
            if enable_streaming_callback_passthrough is not None
            else self.enable_streaming_callback_passthrough
        )

        # Only keep messages with tool calls
        messages_with_tool_calls = [message for message in messages if message.tool_calls]
        streaming_callback = select_streaming_callback(
            init_callback=self.streaming_callback, runtime_callback=streaming_callback, requires_async=True
        )

        if not messages_with_tool_calls:
            return {"tool_messages": [], "state": state}

        # 1) Collect all tool calls and their parameters for parallel execution
        tool_messages = []
        tool_calls, tool_call_params, error_messages = self._prepare_tool_call_params(
            messages_with_tool_calls=messages_with_tool_calls,
            state=state,
            streaming_callback=streaming_callback,
            enable_streaming_passthrough=resolved_enable_streaming_passthrough,
            tools_with_names=tools_with_names,
        )
        tool_messages.extend(error_messages)

        if not tool_call_params:
            return {"tool_messages": tool_messages, "state": state}

        # 2) Execute valid tool calls in parallel
        tool_call_tasks = []
        with ThreadPoolExecutor(max_workers=self.max_workers) as executor:
            for params in tool_call_params:
                loop = asyncio.get_running_loop()
                callable_ = ToolInvoker._make_context_bound_invoke(params["tool_to_invoke"], params["final_args"])
                tool_call_tasks.append(loop.run_in_executor(executor, callable_))

            # 3) Gather and process results: handle errors and merge outputs into state
            tool_results = await asyncio.gather(*tool_call_tasks)
            for tool_result, tool_call in zip(tool_results, tool_calls):
                # a) This is an error, create error Tool message
                if isinstance(tool_result, ToolInvocationError):
                    if self.raise_on_failure:
                        raise tool_result
                    logger.error("{error_exception}", error_exception=tool_result)
                    tool_messages.append(
                        ChatMessage.from_tool(tool_result=str(tool_result), origin=tool_call, error=True)
                    )
                else:
                    # b) In case of success, merge outputs into state
                    try:
                        tool_to_invoke = tools_with_names[tool_call.tool_name]
                        self._merge_tool_outputs(tool=tool_to_invoke, result=tool_result, state=state)
                        tool_messages.append(
                            self._prepare_tool_result_message(
                                result=tool_result, tool_call=tool_call, tool_to_invoke=tool_to_invoke
                            )
                        )
                    except Exception as e:
                        error = ToolOutputMergeError.from_exception(tool_name=tool_call.tool_name, error=e)
                        if self.raise_on_failure:
                            raise error from e
                        logger.error("{error_exception}", error_exception=error)
                        tool_messages.append(
                            ChatMessage.from_tool(tool_result=str(error), origin=tool_call, error=True)
                        )

                # c) Handle streaming callback
                if streaming_callback is not None:
                    await streaming_callback(
                        self._create_tool_result_streaming_chunk(tool_messages=tool_messages, tool_call=tool_call)
                    )

        # 4) We stream one more chunk that contains a finish_reason if tool_messages were generated
        if len(tool_messages) > 0 and streaming_callback is not None:
            await streaming_callback(
                StreamingChunk(
                    content="", finish_reason="tool_call_results", meta={"finish_reason": "tool_call_results"}
                )
            )

        return {"tool_messages": tool_messages, "state": state}

    def to_dict(self) -> dict[str, Any]:
        """
        Serializes the component to a dictionary.

        :returns:
            Dictionary with serialized data.
        """
        if self.streaming_callback is not None:
            streaming_callback = serialize_callable(self.streaming_callback)
        else:
            streaming_callback = None

        return default_to_dict(
            self,
            tools=serialize_tools_or_toolset(self.tools),
            raise_on_failure=self.raise_on_failure,
            convert_result_to_json_string=self.convert_result_to_json_string,
            streaming_callback=streaming_callback,
            enable_streaming_callback_passthrough=self.enable_streaming_callback_passthrough,
            max_workers=self.max_workers,
        )

    @classmethod
    def from_dict(cls, data: dict[str, Any]) -> "ToolInvoker":
        """
        Deserializes the component from a dictionary.

        :param data:
            The dictionary to deserialize from.
        :returns:
            The deserialized component.
        """
        deserialize_tools_or_toolset_inplace(data["init_parameters"], key="tools")
        if data["init_parameters"].get("streaming_callback") is not None:
            data["init_parameters"]["streaming_callback"] = deserialize_callable(
                data["init_parameters"]["streaming_callback"]
            )
        return default_from_dict(cls, data)

deepset-ai / haystack / 18530018322

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous