mirascope 1.25.7__py3-none-any.whl → 2.0.0__py3-none-any.whl

mirascope/llm/tools/decorator.py

@@ -0,0 +1,184 @@
+"""The `llm.tool` decorator for turning functions into tools."""
+
+from dataclasses import dataclass
+from typing import overload
+
+from ..context import DepsT
+from ..types import JsonableCovariantT, P
+from . import _utils as _tool_utils
+from .protocols import AsyncContextToolFn, AsyncToolFn, ContextToolFn, ToolFn
+from .tools import AsyncContextTool, AsyncTool, ContextTool, Tool
+
+
+@dataclass(kw_only=True)
+class ToolDecorator:
+    """Protocol for the tool decorator."""
+
+    strict: bool | None = None
+    """Whether to use strict tool calling, if supported by the provider.
+    
+    If set to None, then it will use the provider's default setting (usually the 
+    strictest possible).
+    """
+
+    @overload
+    def __call__(  # pyright:ignore[reportOverlappingOverload]
+        self, fn: ContextToolFn[DepsT, P, JsonableCovariantT]
+    ) -> ContextTool[DepsT, JsonableCovariantT, P]:
+        """Call the decorator with a context function."""
+        ...
+
+    @overload
+    def __call__(  # pyright:ignore[reportOverlappingOverload]
+        self, fn: AsyncContextToolFn[DepsT, P, JsonableCovariantT]
+    ) -> AsyncContextTool[DepsT, JsonableCovariantT, P]:
+        """Call the decorator with an async context function."""
+        ...
+
+    @overload
+    def __call__(
+        self, fn: ToolFn[P, JsonableCovariantT]
+    ) -> Tool[P, JsonableCovariantT]:
+        """Call the decorator with a sync function."""
+        ...
+
+    @overload
+    def __call__(
+        self, fn: AsyncToolFn[P, JsonableCovariantT]
+    ) -> AsyncTool[P, JsonableCovariantT]:
+        """Call the decorator with an async function."""
+        ...
+
+    def __call__(
+        self,
+        fn: ContextToolFn[DepsT, P, JsonableCovariantT]
+        | AsyncContextToolFn[DepsT, P, JsonableCovariantT]
+        | ToolFn[P, JsonableCovariantT]
+        | AsyncToolFn[P, JsonableCovariantT],
+    ) -> (
+        ContextTool[DepsT, JsonableCovariantT, P]
+        | AsyncContextTool[DepsT, JsonableCovariantT, P]
+        | Tool[P, JsonableCovariantT]
+        | AsyncTool[P, JsonableCovariantT]
+    ):
+        """Call the decorator with a function."""
+        is_context = _tool_utils.is_context_tool_fn(fn)
+        is_async = _tool_utils.is_async_tool_fn(fn)
+
+        if is_context and is_async:
+            return AsyncContextTool[DepsT, JsonableCovariantT, P].from_function(
+                fn, strict=self.strict
+            )
+        elif is_context:
+            return ContextTool[DepsT, JsonableCovariantT, P].from_function(
+                fn, strict=self.strict
+            )
+        elif is_async:
+            return AsyncTool[P, JsonableCovariantT].from_function(
+                fn, strict=self.strict
+            )
+        else:
+            return Tool[P, JsonableCovariantT].from_function(fn, strict=self.strict)
+
+
+@overload
+def tool(  # pyright:ignore[reportOverlappingOverload]
+    __fn: AsyncContextToolFn[DepsT, P, JsonableCovariantT],
+) -> AsyncContextTool[DepsT, JsonableCovariantT, P]:
+    """Overload for async context tool functions."""
+    ...
+
+
+@overload
+def tool(  # pyright:ignore[reportOverlappingOverload]
+    __fn: ContextToolFn[DepsT, P, JsonableCovariantT],
+) -> ContextTool[DepsT, JsonableCovariantT, P]:
+    """Overload for context tool functions."""
+    ...
+
+
+@overload
+def tool(__fn: AsyncToolFn[P, JsonableCovariantT]) -> AsyncTool[P, JsonableCovariantT]:
+    """Overload for regular async tool functions."""
+    ...
+
+
+@overload
+def tool(__fn: ToolFn[P, JsonableCovariantT]) -> Tool[P, JsonableCovariantT]:
+    """Overload for regular sync tool functions."""
+    ...
+
+
+@overload
+def tool(*, strict: bool | None = None) -> ToolDecorator:
+    """Overload for setting non-default arguments."""
+    ...
+
+
+def tool(
+    __fn: ContextToolFn[DepsT, P, JsonableCovariantT]
+    | AsyncContextToolFn[DepsT, P, JsonableCovariantT]
+    | ToolFn[P, JsonableCovariantT]
+    | AsyncToolFn[P, JsonableCovariantT]
+    | None = None,
+    *,
+    strict: bool | None = None,
+) -> (
+    ContextTool[DepsT, JsonableCovariantT, P]
+    | AsyncContextTool[DepsT, JsonableCovariantT, P]
+    | Tool[P, JsonableCovariantT]
+    | AsyncTool[P, JsonableCovariantT]
+    | ToolDecorator
+):
+    '''Decorator that turns a function into a tool definition.
+
+    This decorator creates a `Tool` or `ContextTool` that can be used with `llm.call`.
+    The function's name, docstring, and type hints are used to generate the
+    tool's metadata.
+
+    If the first parameter is named 'ctx' or typed as `llm.Context[T]`, it creates
+    a ContextTool. Otherwise, it creates a regular Tool.
+
+    Args:
+        strict: Whether the tool should use strict mode when supported by the model.
+            If None, uses provider's default (usually as strict as possible).
+
+    Returns:
+        A decorator function that converts the function into a Tool or ContextTool.
+
+    Examples:
+
+        Regular tool:
+        ```python
+        from mirascope import llm
+
+        @llm.tool
+        def available_books() -> list[str]:
+            """Returns the list of available books."""
+            return ["The Name of the Wind"]
+        ```
+
+        Context tool:
+        ```python
+        from dataclasses import dataclass
+
+        from mirascope import llm
+
+
+        @dataclass
+        class Library:
+            books: list[str]
+
+
+        library = Library(books=["Mistborn", "Gödel, Escher, Bach", "Dune"])
+
+        @llm.tool
+        def available_books(ctx: llm.Context[Library]) -> list[str]:
+            """Returns the list of available books."""
+            return ctx.deps.books
+        ```
+    '''
+    decorator = ToolDecorator(strict=strict)
+    if __fn is None:
+        return decorator
+    return decorator(__fn)

mirascope/llm/tools/protocols.py

@@ -0,0 +1,96 @@
+from typing import Any, Protocol
+
+from ..context import Context, DepsT
+from ..types import JsonableCovariantT, P
+
+
+class ToolFn(Protocol[P, JsonableCovariantT]):
+    """Protocol for the tool function."""
+
+    __name__: str
+    """The name of the tool."""
+
+    def __call__(self, *args: P.args, **kwargs: P.kwargs) -> JsonableCovariantT:
+        """Call the function with the given arguments."""
+        raise NotImplementedError()
+
+
+class AsyncToolFn(Protocol[P, JsonableCovariantT]):
+    """Protocol for the async tool function."""
+
+    __name__: str
+    """The name of the tool."""
+
+    async def __call__(self, *args: P.args, **kwargs: P.kwargs) -> JsonableCovariantT:
+        """Call the function with the given arguments."""
+        raise NotImplementedError()
+
+
+class ContextToolFn(Protocol[DepsT, P, JsonableCovariantT]):
+    """Protocol for the context tool function."""
+
+    __name__: str
+    """The name of the tool."""
+
+    def __call__(
+        self, ctx: Context[DepsT], *args: P.args, **kwargs: P.kwargs
+    ) -> JsonableCovariantT:
+        """Call the function with the given arguments."""
+        raise NotImplementedError()
+
+
+class AsyncContextToolFn(Protocol[DepsT, P, JsonableCovariantT]):
+    """Protocol for the async context tool function."""
+
+    __name__: str
+    """The name of the tool."""
+
+    async def __call__(
+        self, ctx: Context[DepsT], *args: P.args, **kwargs: P.kwargs
+    ) -> JsonableCovariantT:
+        """Call the function with the given arguments."""
+        raise NotImplementedError()
+
+
+class KwargsCallable(Protocol[JsonableCovariantT]):
+    """Protocol for functions that can be called with `Any`-typed kwargs.
+
+    Used internally to type-cast tool functions for compatibility with
+    json.loads() output when executing tool calls.
+    """
+
+    def __call__(self, **kwargs: dict[str, Any]) -> JsonableCovariantT: ...
+
+
+class AsyncKwargsCallable(Protocol[JsonableCovariantT]):
+    """Protocol for async functions that can be called with `Any`-typed kwargs.
+
+    Used internally to type-cast async tool functions for compatibility with
+    json.loads() output when executing tool calls.
+    """
+
+    async def __call__(self, **kwargs: dict[str, Any]) -> JsonableCovariantT: ...
+
+
+class ContextKwargsCallable(Protocol[DepsT, JsonableCovariantT]):
+    """Protocol for context functions that can be called with `Any`-typed kwargs.
+
+    Used internally to type-cast context tool functions for compatibility with
+    json.loads() output when executing tool calls.
+    """
+
+    def __call__(
+        self, ctx: Context[DepsT], **kwargs: dict[str, Any]
+    ) -> JsonableCovariantT: ...
+
+
+class AsyncJsonKwargsCallable(Protocol[DepsT, JsonableCovariantT]):
+    """Protocol for async context functions that can be called with `Any`-typed kwargs.
+
+    Used internally to type-cast async context tool functions for compatibility with
+    json.loads() output when executing tool calls.
+    """
+
+    async def __call__(
+        self, ctx: Context[DepsT], **kwargs: dict[str, Any]
+    ) -> JsonableCovariantT: ...

mirascope/llm/tools/tool_schema.py

@@ -0,0 +1,314 @@
+"""The `ToolSchema` class for defining tools that LLMs can request be called."""
+
+from __future__ import annotations
+
+import inspect
+from dataclasses import dataclass
+from typing import (
+    Annotated,
+    Any,
+    Generic,
+    NamedTuple,
+    TypeAlias,
+    TypedDict,
+    TypeVar,
+    cast,
+    get_args,
+    get_origin,
+    get_type_hints,
+)
+
+from docstring_parser import parse
+from pydantic import BaseModel, Field, create_model
+from pydantic.fields import FieldInfo
+
+from ..content import ToolCall
+from ..types import Jsonable
+from .protocols import AsyncContextToolFn, AsyncToolFn, ContextToolFn, ToolFn
+
+AnyToolFn: TypeAlias = (
+    ToolFn[..., Jsonable]
+    | AsyncToolFn[..., Jsonable]
+    | ContextToolFn[Any, ..., Jsonable]
+    | AsyncContextToolFn[Any, ..., Jsonable]
+)
+
+ToolFnT = TypeVar(
+    "ToolFnT",
+    bound=AnyToolFn,
+    covariant=True,
+)
+
+AnyToolSchema: TypeAlias = "ToolSchema[AnyToolFn]"
+ToolSchemaT = TypeVar("ToolSchemaT", bound=AnyToolSchema, covariant=True)
+
+
+ModelJsonSchema = TypedDict(
+    "ModelJsonSchema",
+    {
+        "properties": dict[str, dict[str, Any]],
+        "required": list[str],
+        "$defs": dict[str, dict[str, Any]],
+    },
+    total=False,
+)
+"""Type for Pydantic's model_json_schema() output.
+
+This TypedDict defines the structure of JSON schemas returned by Pydantic models,
+allowing us to avoid type casts when extracting schema components.
+"""
+
+
+class DocstringArg(NamedTuple):
+    """A parameter from a docstring."""
+
+    name: str
+    description: str
+
+
+FORMAT_TOOL_NAME = "__mirascope_formatted_output_tool__"
+"""Reserved name of the formatted output tool. 
+
+Any call to a tool with this name is NOT considered a regular tool call, but will instead
+be converted into textual output containing the arguments to the tool call.
+"""
+# TODO: Investigate whether using this hardcoded tool name has any adverse impact
+# on model performance, compared to a tool name that references the name of the formatted
+# class.
+
+
+@dataclass
+class ParsedDocstring:
+    args: list[DocstringArg]
+
+
+def _parse_docstring_params(docstring: str | None) -> ParsedDocstring:
+    """Parse parameter descriptions from a docstring.
+
+    Uses docstring-parser library which supports ReST, Google, Numpydoc-style and
+    Epydoc docstrings
+
+    Args:
+        docstring: The function's docstring
+
+    Returns:
+        ParsedDocstring containing parameter descriptions
+    """
+    if not docstring:
+        return ParsedDocstring(args=[])
+
+    parsed = parse(docstring)
+    args: list[DocstringArg] = []
+
+    for param in parsed.params:
+        if param.description:
+            args.append(
+                DocstringArg(name=param.arg_name, description=param.description)
+            )
+
+    return ParsedDocstring(args=args)
+
+
+class ToolParameterSchema(BaseModel):
+    """JSON Schema for tool parameters (always an object with properties).
+
+    This contains real JSON Schema as generated by Pydantic, with full support
+    for complex schemas like anyOf, nested objects, validation constraints, etc.
+    Including $defs for complex type references.
+    """
+
+    properties: dict[str, dict[str, Any]] = Field(default_factory=dict)
+    """Dictionary mapping parameter names to their JSON Schema definitions."""
+
+    required: list[str] = Field(default_factory=list)
+    """List of required parameter names."""
+
+    additionalProperties: bool = False
+    """Whether additional properties beyond those defined are allowed."""
+
+    defs: dict[str, dict[str, Any]] | None = Field(default=None, alias="$defs")
+    """JSON Schema definitions for complex types referenced via $ref."""
+
+
+@dataclass
+class ToolSchema(Generic[ToolFnT]):
+    """Underlying schema defining a tool that can be used by LLMs.
+
+    A ToolSchema represents a function that can be called by an LLM during a call.
+    It includes metadata like name, description, and parameter schema.
+
+    This class is not instantiated directly but created by the `@tool()` decorator.
+    """
+
+    fn: ToolFnT
+    """The function that implements the tool's functionality."""
+
+    name: str
+    """The name of the tool, used by the LLM to identify which tool to call."""
+
+    description: str
+    """Description of what the tool does, extracted from the function's docstring."""
+
+    parameters: ToolParameterSchema
+    """JSON Schema describing the parameters accepted by the tool.
+    
+    The serialized parameters table is cached for efficient hash lookups (e.g. when
+    caching provider-encoded tool representations in a LRU cache). Therefore,
+    it should **not be modified** after the ToolSchema is created.
+    """
+
+    strict: bool | None
+    """Whether the tool should use strict mode when supported by the model.
+    
+    If set to None, will use the provider's default setting (usually as strict as 
+    possible).
+    """
+
+    def __hash__(self) -> int:
+        if not hasattr(self, "_hash"):
+            self._hash = hash(
+                (
+                    self.name,
+                    self.description,
+                    self.strict,
+                    self.parameters.model_dump_json(),
+                )
+            )
+        return self._hash
+
+    def __init__(
+        self,
+        fn: ToolFnT,
+        name: str,
+        description: str,
+        parameters: ToolParameterSchema,
+        *,
+        strict: bool | None = None,
+    ) -> None:
+        """Create a `ToolSchema` with the provided values.
+
+        Args:
+            fn: The function that implements the tool's functionality
+            name: The name of the tool
+            description: Description of what the tool does
+            parameters: JSON Schema describing the parameters accepted by the tool
+            strict: Whether the tool should use strict mode when supported.
+                If None, uses provider's default (usually as strict as possible).
+        """
+        self.fn = fn
+        self.name = name
+        self.description = description
+        self.parameters = parameters
+        self.strict = strict
+
+    @classmethod
+    def from_function(
+        cls,
+        fn: AnyToolFn,
+        *,
+        strict: bool | None = None,
+        is_context_tool: bool = False,
+    ) -> ToolSchema[AnyToolFn]:
+        """Create a `ToolSchema` by inspecting a function and its docstring.
+
+        Uses Pydantic's create_model to dynamically build a model from the function
+        signature, then extracts the JSON schema from it. This leverages Pydantic's
+        robust type conversion and union handling.
+
+        Args:
+            fn: The function to extract schema from
+            strict: Whether the tool should use strict mode when supported.
+                If None, uses provider's default (usually as strict as possible).
+            is_context_tool: Whether this is a context tool (skips the context parameter)
+
+        Returns:
+            a `ToolSchema` representing the function
+
+        Raises:
+            ValueError: If the tool has a reserved name.
+        """
+        name = fn.__name__
+        if name.startswith(FORMAT_TOOL_NAME):
+            raise ValueError(
+                f"Cannot use reserved name {FORMAT_TOOL_NAME} as tool name."
+            )
+        description = inspect.cleandoc(fn.__doc__) if fn.__doc__ else name
+
+        param_descriptions = _parse_docstring_params(fn.__doc__)
+
+        field_definitions: dict[str, tuple[Any, Any]] = {}
+        hints = get_type_hints(fn, include_extras=True)
+
+        context_param_skipped = False
+        for param in inspect.signature(fn).parameters.values():
+            # Skip args that are provided by Python or Mirascope, not LLM-generated.
+            # TODO: Handling of FromCallArgs
+            if param.name in ("self", "cls"):
+                continue
+
+            # For context tools, skip the first non-self/cls parameter (the context parameter)
+            if is_context_tool and not context_param_skipped:
+                context_param_skipped = True
+                continue
+
+            param_type = hints.get(param.name, Any)
+            default = ... if param.default is inspect.Parameter.empty else param.default
+            field_info = None
+
+            if get_origin(param_type) is Annotated:
+                args = get_args(param_type)
+                param_type = args[0]
+                for annotation in args[1:]:
+                    if isinstance(annotation, FieldInfo):
+                        field_info = annotation
+                        break
+
+            if field_info is not None:
+                field_value = Field(
+                    default=default,
+                    description=field_info.description,
+                )
+            else:
+                docstring_description: str | None = None
+                for arg in param_descriptions.args:
+                    if arg.name == param.name:
+                        docstring_description = arg.description
+                        break
+                if docstring_description:
+                    field_value = Field(
+                        default=default, description=docstring_description
+                    )
+                else:
+                    field_value = default
+
+            field_definitions[param.name] = (param_type, field_value)
+
+        TempModel = create_model("TempModel", **cast(dict[str, Any], field_definitions))
+
+        schema = cast(ModelJsonSchema, TempModel.model_json_schema())
+
+        parameters = ToolParameterSchema(
+            properties=schema.get("properties", {}),
+            required=schema.get("required", []),
+            additionalProperties=False,
+        )
+        if "$defs" in schema:
+            parameters.defs = schema["$defs"]
+
+        return cls(
+            fn=cast(ToolFnT, fn),
+            name=name,
+            description=description,
+            parameters=parameters,
+            strict=strict,
+        )
+
+    def can_execute(self, tool_call: ToolCall) -> bool:
+        """Check if a `ToolCall` can be executed by tools with this `ToolSchema`.
+
+        This method is a convenient way to determine if a `ToolCall` is likely intended
+        to be executed by a tool with this `ToolSchema`. It does so by checking
+        whether the name on the call matches the name on the schema. No other validation
+        is performed.
+        """
+        return tool_call.name == self.name