- Remove double implementations of functions. The single input is just
taking up space.
- Added tool specific information for `async + showing invoke vs.
ainvoke.
- Added more general information about about `async` (this should live
in a different place eventually since it's not specific to tools).
- Changed ordering of custom tools (StructuredTool is simpler and should
appear before the inheritance)
- Improved the error handling section (not convinced it should be here
though)
"When constructing your own agent, you will need to provide it with a list of Tools that it can use. Besides the actual function that is called, the Tool consists of several components:\n",
"When constructing an agent, you will need to provide it with a list of `Tool`s that it can use. Besides the actual function that is called, the Tool consists of several components:\n",
"\n",
"- `name` (str), is required and must be unique within a set of tools provided to an agent\n",
"- `description` (str), is optional but recommended, as it is used by an agent to determine tool use\n",
"- `args_schema` (Pydantic BaseModel), is optional but recommended, can be used to provide more information (e.g., few-shot examples) or validation for expected parameters.\n",
"| name | str | Must be unique within a set of tools provided to an LLM or agent. |\n",
"| description | str | Describes what the tool does. Used as context by the LLM or agent. |\n",
"| args_schema | Pydantic BaseModel | Optional but recommended, can be used to provide more information (e.g., few-shot examples) or validation for expected parameters |\n",
"| return_direct | boolean | Only relevant for agents. When True, after invoking the given tool, the agent will stop and return the result direcly to the user. |\n",
"\n",
"LangChain provides 3 ways to create tools:\n",
"\n",
"There are multiple ways to define a tool. In this guide, we will walk through how to do for two functions:\n",
"1. Using [@tool decorator](https://api.python.langchain.com/en/latest/tools/langchain_core.tools.tool.html#langchain_core.tools.tool) -- the simplest way to define a custom tool.\n",
"2. Using [StructuredTool.from_function](https://api.python.langchain.com/en/latest/tools/langchain_core.tools.StructuredTool.html#langchain_core.tools.StructuredTool.from_function) class method -- this is similar to the `@tool` decorator, but allows more configuration and specification of both sync and async implementations.\n",
"3. By sub-classing from [BaseTool](https://api.python.langchain.com/en/latest/tools/langchain_core.tools.BaseTool.html) -- This is the most flexible method, it provides the largest degree of control, at the expense of more effort and code.\n",
"\n",
"1. A made up search function that always returns the string \"LangChain\"\n",
"2. A multiplier function that will multiply two numbers by eachother\n",
"The `@tool` or the `StructuredTool.from_function` class method should be sufficient for most use cases.\n",
"\n",
"The biggest difference here is that the first function only requires one input, while the second one requires multiple. Many agents only work with functions that require single inputs, so it's important to know how to work with those. For the most part, defining these custom tools is the same, but there are some differences."
"# Let's inspect some of the attributes associated with the tool.\n",
"print(multiply.name)\n",
"print(multiply.description)\n",
"print(multiply.args)"
"print(multiply.args)\n",
"print(multiply.return_direct)"
]
},
{
"cell_type": "markdown",
"id": "98d6eee9",
"id": "b63fcc3b",
"metadata": {},
"source": [
"You can also customize the tool name and JSON args by passing them into the tool decorator."
"## StructuredTool\n",
"\n",
"The `StrurcturedTool.from_function` class method provides a bit more configurability than the `@tool` decorator, without requiring much additional code."
]
},
{
"cell_type": "code",
"execution_count": 43,
"id": "dbbf4b6c",
"execution_count": 4,
"id": "564fbe6f-11df-402d-b135-ef6ff25e1e63",
"metadata": {},
"outputs": [],
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"6\n",
"10\n"
]
}
],
"source": [
"class SearchInput(BaseModel):\n",
" query: str = Field(description=\"should be a search query\")\n",
"You can also explicitly define a custom tool by subclassing the BaseTool class. This provides maximal control over the tool definition, but is a bit more work."
"You can define a custom tool by sub-classing from `BaseTool`. This provides maximal control over the tool definition, but requires writing more code."
]
},
{
"cell_type": "code",
"execution_count": 45,
"execution_count": 16,
"id": "1dad8f8e",
"metadata": {},
"outputs": [],
@ -188,10 +262,8 @@
" AsyncCallbackManagerForToolRun,\n",
" CallbackManagerForToolRun,\n",
")\n",
"\n",
"\n",
"class SearchInput(BaseModel):\n",
" query: str = Field(description=\"should be a search query\")\n",
"from langchain.pydantic_v1 import BaseModel\n",
"from langchain_core.tools import BaseTool\n",
"\n",
"\n",
"class CalculatorInput(BaseModel):\n",
@ -199,24 +271,6 @@
" b: int = Field(description=\"second number\")\n",
"\n",
"\n",
"class CustomSearchTool(BaseTool):\n",
" name = \"custom_search\"\n",
" description = \"useful for when you need to answer questions about current events\"\n",
"You can also use a `StructuredTool` dataclass. This methods is a mix between the previous two. It's more convenient than inheriting from the BaseTool class, but provides more functionality than just using a decorator."
]
},
{
"cell_type": "code",
"execution_count": 35,
"id": "56ff7670",
"metadata": {
"tags": []
},
"outputs": [],
"source": [
"def search_function(query: str):\n",
" return \"LangChain\"\n",
"LangChain Tools implement the [Runnable interface 🏃](https://api.python.langchain.com/en/latest/runnables/langchain_core.runnables.base.Runnable.html).\n",
"\n",
"All Runnables expose the `invoke` and `ainvoke` methods (as well as other methods like `batch`, `abatch`, `astream` etc).\n",
"\n",
"search = StructuredTool.from_function(\n",
" func=search_function,\n",
" name=\"Search\",\n",
" description=\"useful for when you need to answer questions about current events\",\n",
" # coroutine= ... <- you can specify an async method if desired as well\n",
")"
"So even if you only provide an `sync` implementation of a tool, you could still use the `ainvoke` interface, but there\n",
"are some important things to know:\n",
"\n",
"* LangChain's by default provides an async implementation that assumes that the function is expensive to compute, so it'll delegate execution to another thread.\n",
"* If you're working in an async codebase, you should create async tools rather than sync tools, to avoid incuring a small overhead due to that thread.\n",
"* If you need both sync and async implementations, use `StructuredTool.from_function` or sub-class from `BaseTool`.\n",
"* If implementing both sync and async, and the sync code is fast to run, override the default LangChain async implementation and simply call the sync code.\n",
"* You CANNOT and SHOULD NOT use the sync `invoke` with an `async` tool."
]
},
{
"cell_type": "code",
"execution_count": 38,
"id": "d3fd3896",
"execution_count": 8,
"id": "6615cb77-fd4c-4676-8965-f92cc71d4944",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Search\n",
"Search(query: str) - useful for when you need to answer questions about current events\n",
"Raised not implemented error. You should not be doing this.\n"
]
}
],
"source": [
"print(calculator.name)\n",
"print(calculator.description)\n",
"print(calculator.args)"
"@tool\n",
"async def multiply(a: int, b: int) -> int:\n",
" \"\"\"Multiply two numbers.\"\"\"\n",
" return a * b\n",
"\n",
"\n",
"try:\n",
" multiply.invoke({\"a\": 2, \"b\": 3})\n",
"except NotImplementedError:\n",
" print(\"Raised not implemented error. You should not be doing this.\")"
]
},
{
"cell_type": "markdown",
"id": "f1da459d",
"id": "f9c746a7-88d7-4afb-bcb8-0e98b891e8b6",
"metadata": {},
"source": [
"## Handling Tool Errors \n",
"When a tool encounters an error and the exception is not caught, the agent will stop executing. If you want the agent to continue execution, you can raise a `ToolException` and set `handle_tool_error` accordingly. \n",
"\n",
"When `ToolException` is thrown, the agent will not stop working, but will handle the exception according to the `handle_tool_error` variable of the tool, and the processing result will be returned to the agent as observation, and printed in red.\n",
"If you're using tools with agents, you will likely need an error handling strategy, so the agent can recover from the error and continue execution.\n",
"\n",
"A simple strategy is to throw a `ToolException` from inside the tool and specify an error handler using `handle_tool_error`. \n",
"\n",
"You can set `handle_tool_error` to `True`, set it a unified string value, or set it as a function. If it's set as a function, the function should take a `ToolException` as a parameter and return a `str` value.\n",
"When the error handler is specified, the exception will be caught and the error handler will decide which output to return from the tool.\n",
"\n",
"You can set `handle_tool_error` to `True`, a string value, or a function. If it's a function, the function should take a `ToolException` as a parameter and return a value.\n",
"\n",
"Please note that only raising a `ToolException` won't be effective. You need to first set the `handle_tool_error` of the tool because its default value is `False`."
"Cell \u001b[0;32mIn[55], line 5\u001b[0m, in \u001b[0;36msearch_tool1\u001b[0;34m(s)\u001b[0m\n\u001b[1;32m 4\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21msearch_tool1\u001b[39m(s: \u001b[38;5;28mstr\u001b[39m):\n\u001b[0;32m----> 5\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m ToolException(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mThe search tool1 is not available.\u001b[39m\u001b[38;5;124m\"\u001b[39m)\n",
"\u001b[0;31mToolException\u001b[0m: The search tool1 is not available."
]
"data": {
"text/plain": [
"'Error: There is no city by the name of foobar.'"
Eugene Yurtsev
4 weeks ago
committed by
GitHub