Copy

docs/docs/how_to/index.mdx

@@ -196,6 +196,7 @@ LangChain [Tools](/docs/concepts/#tools) contain a description of the tool (to p
 - [How to: handle errors when calling tools](/docs/how_to/tools_error)
 - [How to: disable parallel tool calling](/docs/how_to/tool_choice)
 - [How to: stream events from within a tool](/docs/how_to/tool_stream_events)
+- [How to: pass additional parameters to tools](/docs/how_to/tools_additional_params)
 
 ### Multimodal
 

docs/docs/how_to/tools_additional_params.ipynb

@@ -0,0 +1,215 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# How to pass additional parameters to tools\n",
+    "\n",
+    ":::info Prerequisites\n",
+    "\n",
+    "This guide assumes familiarity with the following concepts:\n",
+    "- [Chat models](/docs/concepts/#chat-models)\n",
+    "- [LangChain Tools](/docs/concepts/#tools)\n",
+    "- [How to use a model to call tools](/docs/how_to/tool_calling)\n",
+    "- [Chaining](/docs/how_to/sequence/)\n",
+    "\n",
+    ":::\n",
+    "\n",
+    "When using a chat model to call a tool in a chain or agent, it can be convenient for that tool to use information aside from arguments generated by the model.\n",
+    "\n",
+    "This guide covers how to use a technique called [currying](https://en.wikipedia.org/wiki/Currying) to predefine certain parameters passed to a tool while letting the model generate the rest. It's similar in spirit to [partially applying variables to a prompt](/docs/how_to/prompts_partial/).\n",
+    "\n",
+    "Let's mock out a function called `weather` as follows. Note that we aren't wrapping it with a `@tool` decorator:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def current_weather(location: str, use_cached_results: bool):\n",
+    "    \"\"\"Use this to get the current weather in a given location.\"\"\"\n",
+    "    if use_cached_results:\n",
+    "        # Use a cached response\n",
+    "        return f\"The current weather in {location} is 50 degrees and rainy (cached).\"\n",
+    "    else:\n",
+    "        # Use a freshly retrieved response\n",
+    "        return f\"The current weather in {location} is 0 degrees and sunny.\""
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "You probably don't want the model to determine our caching strategy by populating `use_cached_results` - instead you probably want to define it yourself based on something like recency.\n",
+    "\n",
+    "To do this, define a `curry` function that accepts a standard Python function, binding arguments to it and converting it into a tool:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import inspect\n",
+    "from functools import wraps\n",
+    "\n",
+    "from langchain_core.tools import StructuredTool\n",
+    "\n",
+    "\n",
+    "def curry(func, **curried_kwargs):\n",
+    "    \"\"\"Util to return partialed functions whose signatures also hide the curried variables.\"\"\"\n",
+    "\n",
+    "    @wraps(func)\n",
+    "    def wrapper(*args, **kwargs):\n",
+    "        # Merge curried kwargs with the new kwargs\n",
+    "        new_kwargs = {**curried_kwargs, **kwargs}\n",
+    "        return func(*args, **new_kwargs)\n",
+    "\n",
+    "    # Get the original function's signature\n",
+    "    sig = inspect.signature(func)\n",
+    "\n",
+    "    # Create a new signature without the curried parameters\n",
+    "    new_params = [p for name, p in sig.parameters.items() if name not in curried_kwargs]\n",
+    "    wrapper.__signature__ = sig.replace(parameters=new_params)\n",
+    "\n",
+    "    return StructuredTool.from_function(wrapper)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "You can wrap the `current_weather` function in the `curry` function to automatically set a `use_cached_result` to `True` and ensure that the model only generates a `location` parameter. Here's what that looks like when we bind it to a chat model:\n",
+    "\n",
+    "```{=mdx}\n",
+    "import ChatModelTabs from \"@theme/ChatModelTabs\";\n",
+    "\n",
+    "<ChatModelTabs customVarName=\"model\" />\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# | output: false\n",
+    "# | echo: false\n",
+    "\n",
+    "%pip install -qU langchain langchain_anthropic\n",
+    "\n",
+    "import os\n",
+    "from getpass import getpass\n",
+    "\n",
+    "from langchain_anthropic import ChatAnthropic\n",
+    "\n",
+    "if \"ANTHROPIC_API_KEY\" not in os.environ:\n",
+    "    os.environ[\"ANTHROPIC_API_KEY\"] = getpass()\n",
+    "\n",
+    "model = ChatAnthropic(model=\"claude-3-5-sonnet-20240620\", temperature=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'name': 'current_weather',\n",
+       "  'args': {'location': 'San Francisco'},\n",
+       "  'id': 'toolu_01KaLhdKqg8CvwCUS7Mmiom6'}]"
+      ]
+     },
+     "execution_count": 9,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "curried_tool = curry(current_weather, use_cached_results=True)\n",
+    "\n",
+    "llm_with_tools = model.bind_tools([curried_tool])\n",
+    "\n",
+    "res = llm_with_tools.invoke(\"What is the weather in SF?\")\n",
+    "\n",
+    "res.tool_calls"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "You can see that the model does not generate a parameter for `use_cached_results`. If you then extend the chain to invoke the tool with the provided arguments, you can see that you'll always get a cached result rather than a fresh one:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'The current weather in San Francisco is 50 degrees and rainy (cached).'"
+      ]
+     },
+     "execution_count": 10,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# Simplified way to call the tool from model output\n",
+    "# Assumes the model always calls the weather tool\n",
+    "chain = llm_with_tools | (lambda x: x.tool_calls[0][\"args\"]) | curried_tool\n",
+    "\n",
+    "chain.invoke(\"What is the weather in SF?\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Next steps\n",
+    "\n",
+    "You've now seen how to pass additional arguments to tools. Next, check out the following guides to learn more about tools:\n",
+    "\n",
+    "- [Build agents using LangGraph](https://langchain-ai.github.io/langgraph/tutorials/introduction/)\n",
+    "- [Force a tool call](/docs/how_to/tool_choice)\n",
+    "- [Pass tool results back to model](/docs/how_to/tool_results_pass_to_model)\n",
+    "\n",
+    "You can also check out the full list of how-to guides on tools:\n",
+    "\n",
+    "- Building [tool-using chains and agents](/docs/how_to#tools)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.5"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}