Langchain学习 - Prompt和Parser

2023-07-16

随着ChatGPT的发布，AI，LLM，GPT领域成为了热点。而Langchain这个库也吸引了越来越多人的目光，Star数量水涨船高。

刚好DeepLearning最近发布了Langchain课程，不过课程内用的是Python，对前端开发者不是很友好。所以我计划用JS实现一遍课程内的内容，记录遇到的问题，算是学习笔记了。废话不多说，第一讲开始

案例

假设你收到一份客户发来的邮件，邮件可能是用其他语言写的。

const customEmail = `
  Arrr, I be fuming that me blender lid
  flew off and splattered me kitchen walls
  with smoothie! And to make matters worse,
  the warranty don't cover the cost of
  cleaning up me kitchen. I need yer help
  right now, matey!
`

我们可以用OpenAI将其翻译成美式英语，以一种平静且尊重的语气

const style = `
  American English
  in a calm and respectful tone
`

我们可以用下面的prompt和OpenAI打交道

const prompt = `
  Translate the text
  that is delimited by double quote
  into a style that is ${style}.
  text: "${customEmail}"
`

最后我们来调用OpenAI API来完成翻译任务。我们先用 openai 这个库来实现

import OpenAI from "openai"; // openai v4
import { HttpsProxyAgent } from "https-proxy-agent";
import "dotenv/config";
// 在国内，需要加上翻墙代理才能调用成功
const agent = new HttpsProxyAgent("http://localhost:8001");
export const openai = new OpenAI({
  apiKey: process.env["OPENAI_API_KEY"],
  httpAgent: agent,
});
async function main() {
  const completion = await openai.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: [{ role: "user", content: prompt }],
    temperature: 0,
  });
  console.log(completion.choices[0].message);
}
main()

来看一下用Langchain实现的话要怎么做

import { ChatOpenAI } from "langchain/chat_models/openai";
import { HttpsProxyAgent } from 'https-proxy-agent'
import 'dotenv/config';
const chat = new ChatOpenAI({ temperature: 0, openAIApiKey: process.env.OPENAI_API_KEY }, {
  baseOptions: {
    httpsAgent: new HttpsProxyAgent('http://localhost:8001')
  }
});

相比于我们之前用模板字符串的方式来做prompt的拼接，Langchain提供了对promp模板的封装

import { ChatPromptTemplate, HumanMessagePromptTemplate } from 'langchain/prompts'
const templateString = `
  Translate the text
  that is delimited by double quote
  into a style that is {style}.
  text: "{text}"
`
const chatPrompt = ChatPromptTemplate.fromPromptMessages([
  HumanMessagePromptTemplate.fromTemplate(templateString)
])

我们可以直接用prompt模板来生成格式化的messages，用于和OpenAI的chat打交道

const customEmail = `
  Arrr, I be fuming that me blender lid
  flew off and splattered me kitchen walls
  with smoothie! And to make matters worse,
  the warranty don't cover the cost of
  cleaning up me kitchen. I need yer help
  right now, matey!
`
const style = `
  American English
  in a calm and respectful tone
`
const messages = await chatPrompt.formatMessages({ style, text: customEmail });
const response = chat.call(messages);
console.log(response.content);

对比一下我们可以发现，Langchain相比原来的写法更简洁和方便。而且对于prompt模板的封装也更利于prompt的抽象和复用。

假设我们需要将其翻译成其他的风格和语言，我们只需要用不同的 style 调用 chatPrompt.formatMessages 就可以

const messages = await chatPrompt.formatMessages({
  style: `a polite tone that speaks in English Pirate`,
  text: customEmail
});

Parser

还有些场景我们期望LLM以特定的格式输出内容，比如JSON。

{
  "gift": false,
  "delivery_days": 5,
  "price_value": "pretty affordable!"
}

假设我们要把产品评价，利用LLM格式化成上面的格式，我们可能会这么实现

const customeReview = `
  This leaf blower is pretty amazing.  It has four settings:
  candle blower, gentle breeze, windy city, and tornado.
  It arrived in two days, just in time for my wife's
  anniversary present.
  I think my wife liked it so much she was speechless.
  So far I've been the only one using it, and I've been
  using it every other morning to clear the leaves on our lawn.
  It's slightly more expensive than the other leaf blowers
  out there, but I think it's worth it for the extra features.
`
const reviewTemplate = `
  For the following text, extract the following information:
  gift: Was the item purchased as a gift for someone else?
  Answer True if yes, False if not or unknown.
  delivery_days: How many days did it take for the product
  to arrive? If this information is not found, output -1.
  price_value: Extract any sentences about the value or price,
  and output them as a comma separated Array.
  Format the output as JSON with the following keys:
  gift
  delivery_days
  price_value
  text: {text}
`
const chatPrompt = ChatPromptTemplate.fromPromptMessages([
  HumanMessagePromptTemplate.fromTemplate(reviewTemplate)
])
const messages = await chatPrompt.formatMessages({ text: customeReview });
const response = await chat.call(messages);
console.log(response.content);

输出内容看起来是个JSON，但其实是一个JSON string。想要得到真正的JSON对象，我们可以用Langchain提供的Output Parser来做

import { StructuredOutputParser } from 'langchain/output_parsers'
const reviewTemplate = `
  For the following text, extract the following information:
  gift: Was the item purchased as a gift for someone else? \
  Answer True if yes, False if not or unknown.
  delivery_days: How many days did it take for the product \
  to arrive? If this information is not found, output -1.
  price_value: Extract any sentences about the value or price,\
  and output them as an Array.
  Format the output as JSON with the following keys:
  gift
  delivery_days
  price_value
  text: {text}
  {format_instructions}
`
const const parser = StructuredOutputParser.fromNamesAndDescriptions({
  gift: 'Was the item purchased as a gift for someone else? Answer True if yes, False if not or unknown.',
  delivery_days: 'How many days did it take for the product to arrive? If this information is not found, output -1.',
  price_value: 'Extract any sentences about the value or price, and output them as an Array.'
})
const format_instructions = parser.getFormatInstructions();
const messages = await chatPrompt.formatMessages({ text: customeReview, format_instructions });
const response = await chat.call(messages);
const output = await parser.parse(response.content);
console.log(output)

我们先在 reviewTemplate 底部加上了 format_insturctions 变量。然后用 StructuredOutputParser 生成 format_insturctions

最后用 parser.parse 来解析OpenAI的输出结果，得到我们想要的格式

不过运行上面的代码是会报错的

ZodError: [
  {
    "code": "invalid_type",
    "expected": "string",
    "received": "boolean",
    "path": [
      "gift"
    ],
    "message": "Expected string, received boolean"
  },
  {
    "code": "invalid_type",
    "expected": "string",
    "received": "number",
    "path": [
      "delivery_days"
    ],
    "message": "Expected string, received number"
  },
  {
    "code": "invalid_type",
    "expected": "string",
    "received": "array",
    "path": [
      "price_value"
    ],
    "message": "Expected string, received array"
  }
]

从报错信息可以是类型不匹配导致的。我们需要用 zod 来定义类型

import { z } from 'zod'
const parser = StructuredOutputParser.fromZodSchema(z.object({
  gift: z.boolean().describe(`Was the item purchased as a gift for someone else? Answer True if yes, False if not or unknown.`),
  delivery_days: z.string().describe(`How many days did it take for the product to arrive? If this information is not found, output -1.`),
  price_value: z.array(z.string()).describe(`Extract any sentences about the value or price, and output them as a comma separated Array.`)
}))

将 parser 改成调用 StructureOutputParser.fromZodSchema 方法，并用 zod 来定义类型，就可以运行成功了

将输出格式化后，我们就可以讲其作为参数调用其他LLM，形成一个调用链，构建更强大复杂的LLM应用