Skip to main content

XML Agent

info

Looking for the non LCEL version of this chain? Click here to view the legacy doc.

Some language models (like Anthropic's Claude) are particularly good at reasoning/writing XML. The below example shows how to use an agent that uses XML when prompting.

import { ChatAnthropic } from "langchain/chat_models/anthropic";
import { AgentExecutor } from "langchain/agents";
import { SerpAPI, Tool } from "langchain/tools";
import {
  ChatPromptTemplate,
  HumanMessagePromptTemplate,
  MessagesPlaceholder,
} from "langchain/prompts";
import { RunnableSequence } from "langchain/schema/runnable";
import { AgentStep } from "langchain/schema";
import { XMLAgentOutputParser } from "langchain/agents/xml/output_parser";
import { renderTextDescription } from "langchain/tools/render";
import { formatLogToMessage } from "langchain/agents/format_scratchpad/log_to_message";

/**
 * Define your chat model.
 * In this case we'll use Claude since it preforms well on XML related tasks
 */
const model = new ChatAnthropic({ modelName: "claude-2", temperature: 0 }).bind(
  {
    stop: ["</tool_input>", "</final_answer>"],
  }
);
/** Define your list of tools. */
const tools = [new SerpAPI()];

/**
 * Construct your prompt.
 * For XML not too much work is necessary, we just need to
 * define our prompt, and a messages placeholder for the
 * previous agent steps.
 */
const AGENT_INSTRUCTIONS = `You are a helpful assistant. Help the user answer any questions.

You have access to the following tools:

{tools}

In order to use a tool, you can use <tool></tool> and <tool_input></tool_input> tags.
You will then get back a response in the form <observation></observation>
For example, if you have a tool called 'search' that could run a google search, in order to search for the weather in SF you would respond:

<tool>search</tool><tool_input>weather in SF</tool_input>
<observation>64 degrees</observation>

When you are done, respond with a final answer between <final_answer></final_answer>. For example:

<final_answer>The weather in SF is 64 degrees</final_answer>

Begin!

Question: {input}`;
const prompt = ChatPromptTemplate.fromMessages([
  HumanMessagePromptTemplate.fromTemplate(AGENT_INSTRUCTIONS),
  new MessagesPlaceholder("agent_scratchpad"),
]);

/**
 * Next construct your runnable agent using a `RunnableSequence`
 * which takes in two arguments: input and agent_scratchpad.
 * The agent_scratchpad is then formatted using the `formatLogToMessage`
 * util because we're using a `MessagesPlaceholder` in our prompt.
 *
 * We also need to pass our tools through formatted as a string since
 * our prompt function does not format the prompt.
 */
const runnableAgent = RunnableSequence.from([
  {
    input: (i: { input: string; tools: Tool[]; steps: AgentStep[] }) => i.input,
    agent_scratchpad: (i: {
      input: string;
      tools: Tool[];
      steps: AgentStep[];
    }) => formatLogToMessage(i.steps),
    tools: (i: { input: string; tools: Tool[]; steps: AgentStep[] }) =>
      renderTextDescription(i.tools),
  },
  prompt,
  model,
  new XMLAgentOutputParser(),
]);

/**
 * Finally, we can define our agent executor and call it with an input.
 */
const executor = AgentExecutor.fromAgentAndTools({
  agent: runnableAgent,
  tools,
});

console.log("Loaded agent.");

const input = `What is the weather in Honolulu?`;
console.log(`Calling executor with input: ${input}`);
const result = await executor.invoke({ input, tools });
console.log(result);

/*
Loaded agent.
Calling executor with input: What is the weather in Honolulu?
{
  output: '\n' +
    'The weather in Honolulu is mostly sunny with a high of 72 degrees Fahrenheit, 2% chance of rain, 91% humidity, and winds around 2 mph.\n'
}
*/

API Reference: