Custom Tools

Overview

Inspect natively supports registering Python functions as tools and providing these tools to models that support them. Inspect also supports secure sandboxes for running abritrary code produced by models, flexible error handling, as well as dynamic tool definitions.

We’ll cover all of these features below, but we’ll start with a very simple example to cover the basic mechanics of tool use.

Defining Tools

Here’s a simple tool that adds two numbers. The @tool decorator is used to register it with the system:

from inspect_ai.tool import tool

@tool
def add():
    async def execute(x: int, y: int):
        """
        Add two numbers.

        Args:
            x: First number to add.
            y: Second number to add.

        Returns:
            The sum of the two numbers.
        """
        return x + y

    return execute

Annotations

Note that we provide type annotations for both arguments:

async def execute(x: int, y: int)

Further, we provide descriptions for each parameter in the documention comment:

Args:
    x: First number to add.
    y: Second number to add.

Type annotations and descriptions are required for tool declarations so that the model can be informed which types to pass back to the tool function and what the purpose of each parameter is.

Note that you while you are required to provide default descriptions for tools and their parameters within doc comments, you can also make these dynamically customisable by users of your tool (see the section on Tool Descriptions for details on how to do this).

Using Tools

We can use the addition() tool in an evaluation by passing it to the use_tools() Solver:

from inspect_ai import Task, task
from inspect_ai.dataset ipmort Sample
from inspect_ai.solver import generate, use_tools
from inspect_ai.scorer import match

@task
def addition_problem():
    return Task(
        dataset=[Sample(input="What is 1 + 1?", target=["2"])],
        solver=[
            use_tools(add()), 
            generate()
        ],
        scorer=match(numeric=True),
    )

Note that this tool doesn’t make network requests or do heavy computation, so is fine to run as inline Python code. If your tool does do more elaborate things, you’ll want to make sure it plays well with Inspect’s concurrency scheme. For network requests, this amounts to using async HTTP calls with httpx. For heavier computation, tools should use subprocesses as described in the next section.

Note that when using tools with models, the models do not call the Python function directly. Rather, the model generates a structured request which includes function parameters, and then Inspect calls the function and returns the result to the model.

Tool Errors

Various errors can occur during tool execution, especially when interacting with the file system or network or when using Sandbox Environments to execute code in a container sandbox. As a tool writer you need to decide how you’d like to handle error conditions. A number of approaches are possible:

1. Notify the model that an error occurred to see whether it can recover.

2. Catch and handle the error internally (trying another code path, etc.).

3. Allow the error to propagate, resulting in the current Sample failing with an error state.

There are no universally correct approaches as tool usage and semantics can vary widely—some rough guidelines are provided below.

Default Handling

If you do not explicitly handle errors, then Inspect provides some default error handling behaviour. Specifically, if any of the following errors are raised they will be handled and reported to the model:

• TimeoutError — Occurs when a call to subprocess() or sandbox().exec() times out.

• PermissionError — Occurs when there are inadequate permissions to read or write a file.

• UnicodeDecodeError — Occurs when the output from executing a process or reading a file is binary rather than text.

• OutputLimitExceededError - Occurs when one or both of the output streams from sandbox().exec() exceed 10 MiB or when attempting to read a file over 100 MiB in size.

• ToolError — Special error thrown by tools to indicate they’d like to report an error to the model.

These are all errors that are expected (in fact the SandboxEnvironment interface documents them as such) and possibly recoverable by the model (try a different command, read a different file, etc.). Unexpected errors (e.g. a network error communicating with a remote service or container runtime) on the other hand are not automatically handled and result in the Sample failing with an error.

Many tools can simply rely on the default handling to provide reasonable behaviour around both expected and unexpected errors.

When we say that the errors are reported directly to the model, this refers to the behaviour when using the default generate(). If on the other hand, you are have created custom scaffolding for an agent, you can intercept tool errors and apply additional filtering and logic.

Explicit Handling

In some cases a tool can implement a recovery strategy for error conditions. For example, an HTTP request might fail due to transient network issues, and retrying the request (perhaps after a delay) may resolve the problem. Explicit error handling strategies are generally applied when there are expected errors that are not already handled by Inspect’s Default Handling.

Another type of explicit handling is re-raising an error to bypass Inspect’s default handling. For example, here we catch at re-raise TimeoutError so that it fails the Sample:

try:
  result = await sandbox().exec(
    cmd=["decode", file], 
    timeout=timeout
  )
except TimeoutError:
  raise RuntimeError("Decode operation timed out.")
  

Sandboxing

Tools may have a need to interact with a sandboxed environment (e.g. to provide models with the ability to execute arbitrary bash or python commands). The active sandbox environment can be obtained via the sandbox() function. For example:

from inspect_ai.tool import ToolError, tool
from inspect_ai.util import sandbox

@tool
def list_files():
    async def execute(dir: str):
        """List the files in a directory.

        Args:
            dir (str): Directory

        Returns:
            File listing of the directory
        """
        result = await sandbox().exec(["ls", dir])
        if result.success:
            return result.stdout
        else:
            raise ToolError(result.stderr)

    return execute

The following instance methods are available to tools that need to interact with a SandboxEnvironment:

class SandboxEnvironment:
   
    async def exec(
        self,
        cmd: list[str],
        input: str | bytes | None = None,
        cwd: str | None = None,
        env: dict[str, str] = {},
        user: str | None = None,
        timeout: int | None = None,
        timeout_retry: bool = True
    ) -> ExecResult[str]:
        """
        Raises:
          TimeoutError: If the specified `timeout` expires.
          UnicodeDecodeError: If an error occurs while
            decoding the command output.
          PermissionError: If the user does not have
            permission to execute the command.
          OutputLimitExceededError: If an output stream
            exceeds the 10 MiB limit.
        """
        ...

    async def write_file(
        self, file: str, contents: str | bytes
    ) -> None:
        """
        Raises:
          PermissionError: If the user does not have
            permission to write to the specified path.
          IsADirectoryError: If the file exists already and 
            is a directory.
        """
        ...

    async def read_file(
        self, file: str, text: bool = True
    ) -> Union[str | bytes]:
        """
        Raises:
          FileNotFoundError: If the file does not exist.
          UnicodeDecodeError: If an encoding error occurs 
            while reading the file.
            (only applicable when `text = True`)
          PermissionError: If the user does not have
            permission to read from the specified path.
          IsADirectoryError: If the file is a directory.
          OutputLimitExceededError: If the file size
            exceeds the 100 MiB limit.
        """
        ...

    async def connection(self) -> SandboxConnection:
        """
        Raises:
           NotImplementedError: For sandboxes that don't provide connections
           ConnectionError: If sandbox is not currently running.
        """

The read_file() method should preserve newline constructs (e.g. crlf should be preserved not converted to lf). This is equivalent to specifying newline="" in a call to the Python open() function. Note that write_file() automatically creates parent directories as required if they don’t exist.

The connection() method is optional, and provides commands that can be used to login to the sandbox container from a terminal or IDE.

Note that to deal with potential unreliability of container services, the exec() method includes a timeout_retry parameter that defaults to True. For sandbox implementations this parameter is advisory (they should only use it if potential unreliablity exists in their runtime). No more than 2 retries should be attempted and both with timeouts less than 60 seconds. If you are executing commands that are not idempotent (i.e. the side effects of a failed first attempt may affect the results of subsequent attempts) then you can specify timeout_retry=False to override this behavior.

For each method there is a documented set of errors that are raised: these are expected errors and can either be caught by tools or allowed to propagate in which case they will be reported to the model for potential recovery. In addition, unexpected errors may occur (e.g. a networking error connecting to a remote container): these errors are not reported to the model and fail the Sample with an error state.

See the documentation on Sandbox Environments for additional details.

Stateful Tools

Some tools need to retain state across invocations (for example, the bash_session() and web_browser() tools both interact with a stateful remote process). You can create stateful tools by using the store_as() function to access discrete storage for your tool and/or speicfic instances of your tool.

For example, imagine we were creating a web_surfer() tool that builds on the web_browser() tool to complete sequences of browser actions in service of researching a topic. We might want to ask multiple questions of the web surfer and have it retain its message history and browser state.

Here’s the complete source code for this tool.

from textwrap import dedent

from pydantic import Field
from shortuuid import uuid

from inspect_ai.model import (
  ChatMessage, ChatMessageSystem, ChatMessageUser, get_model
)
from inspect_ai.tool import Tool, tool, web_browser
from inspect_ai.util import StoreModel, store_as

class WebSurferState(StoreModel):
    messages: list[ChatMessage] = Field(default_factory=list)

@tool
def web_surfer(instance: str | None = uuid()) -> Tool:
    """Web surfer tool for researching topics.

    The web_surfer tool builds on the web_browser tool to complete
    sequences of web_browser actions in service of researching a topic.
    Input can either be requests to do research or questions about 
    previous research.
    """

    async def execute(input: str, clear_history: bool = False) -> str:
        """Use the web to research a topic.

        You may ask the web surfer any question. These questions can 
        either prompt new web searches or be clarifying or follow up 
        questions about previous web searches.

        Args:
           input: Message to the web surfer. This can be a prompt to
              do research or a question about previous research.
           clear_history: Clear memory of previous searches.

        Returns:
           Answer to research prompt or question.
        """
        # keep track of message history in the store
        surfer_state = store_as(WebSurferState, instance=instance)

        # clear history if requested.
        if clear_history:
            surfer_state.messages.clear()

        # provide system prompt if we are at the beginning
        if len(surfer_state.messages) == 0:
            surfer_state.messages.append(
                ChatMessageSystem(
                    content=dedent("""
                You are a helpful assistant that can use a browser
                to answer questions. You don't need to answer the 
                questions with a single web browser request, rather,
                you can perform searches, follow links, backtrack, 
                and otherwise use the browser to its fullest 
                capability to help answer the question.

                In some cases questions will be about your previous
                web searches, in those cases you don't always need
                to use the web browser tool but can answer by 
                consulting previous conversation messages.
                """)
                )
            )

        # append the latest question
        surfer_state.messages.append(ChatMessageUser(content=input))

        # run tool loop with web browser
        messages, output = await get_model().generate_loop(
            surfer_state.messages, tools=web_browser(instance=instance)
        )

        # update state
        surfer_state.messages.extend(messages)

        # return response
        return output.completion

    return execute

Note that we use an instance parameter (which defaults to uuid()) to ensure that each call to the web_surfer() function creates its own instance of the tool. We then pass this instance to the store_as() function (to store our own tool’s message history) and the web_browser() function (so that we also provision a unique browser for the web surfer session).

Tool Choice

By default models will use a tool if they think it’s appropriate for the given task. You can override this behaviour using the tool_choice parameter of the use_tools() Solver. For example:

# let the model decide whether to use the tool
use_tools(addition(), tool_choice="auto")

# force the use of a tool
use_tools(addition(), tool_choice=ToolFunction(name="addition"))

# prevent use of tools
use_tools(addition(), tool_choice="none")

The last form (tool_choice="none") would typically be used to turn off tool usage after an initial generation where the tool used. For example:

solver = [
  use_tools(addition(), tool_choice=ToolFunction(name="addition")),
  generate(),
  follow_up_prompt(),
  use_tools(tool_choice="none"),
  generate()
]

Tool Descriptions

Well crafted tools should include descriptions that provide models with the context required to use them correctly and productively. If you will be developing custom tools it’s worth taking some time to learn how to provide good tool definitions. Here are some resources you may find helpful:

• Best Practices for Tool Definitions
• Function Calling with LLMs

In some cases you may want to change the default descriptions created by a tool author—for example you might want to provide better disambiguation between multiple similar tools that are used together. You also might have need to do this during development of tools (to explore what descriptions are most useful to models).

The tool_with() function enables you to take any tool and adapt its name and/or descriptions. For example:

from inspect_ai.tool import tool_with

my_add = tool_with(
  tool=add(), 
  name="my_add",
  description="a tool to add numbers", 
  parameters={
    "x": "the x argument",
    "y": "the y argument"
  })

You need not provide all of the parameters shown above, for example here are some examples where we modify just the main tool description or only a single parameter:

my_add1 = tool_with(add(), description="a tool to add numbers")
my_add2 = tool_with(add(), parameters={"x": "the x argument"})

Note that tool_with() function modifies the passed tool in-place, so if you want to create multiple variations of a single tool using tool_with() you should create the underlying tool multiple times, once for each call to tool_with() (this is demonsrated in the example above).

Dynamic Tools

As described above, normally tools are defined using @tool decorators and documentation comments. It’s also possible to create a tool dynamically from any function by creating a ToolDef. For example:

from inspect_ai.solver import use_tools
from inspect_ai.tool import ToolDef

async def addition(x: int, y: int):
    return x + y

add = ToolDef(
    tool=addition,
    name="add",
    description="A tool to add numbers", 
    parameters={
        "x": "the x argument",
        "y": "the y argument"
    })
)

use_tools([add])

This is effectively what happens under the hood when you use the @tool decorator. There is one critical requirement for functions that are bound to tools using ToolDef: type annotations must be provided in the function signature (e.g. x: int, y: int).

For Inspect APIs, ToolDef can generally be used anywhere that Tool can be used (use_tools(), setting state.tools, etc.). If you are using a 3rd party API that does not take Tool in its interface, use the ToolDef.as_tool() method to adapt it. For example:

from inspect_agents import my_agent
agent = my_agent(tools=[add.as_tool()])

If on the other hand you want to get the ToolDef for an existing tool (e.g. to discover its name, description, and parameters) you can just pass the Tool to the ToolDef constructor (including whatever overrides for name, etc. you want):

from inspect_ai.tool import ToolDef, bash
bash_def = ToolDef(bash())