106 lines
2.8 KiB
TypeScript
106 lines
2.8 KiB
TypeScript
import { PromptTemplate } from "@langchain/core/prompts"
|
||
import { z } from "zod"
|
||
|
||
import { getLangfuse, getModel } from "./base"
|
||
|
||
const groupAgentConfig = z.object({
|
||
chatId: z.string().describe("群聊ID"),
|
||
chatName: z.string().describe("群聊名称"),
|
||
functionId: z.string().describe("功能ID"),
|
||
functionName: z.string().describe("功能名称"),
|
||
startTime: z.string().describe("开始时间，格式为 YYYY-MM-DD HH:mm:ss"),
|
||
endTime: z.string().describe("结束时间，格式为 YYYY-MM-DD HH:mm:ss"),
|
||
})
|
||
|
||
/**
|
||
* 解析GroupAgent用户输入
|
||
* @param userInput 用户输入
|
||
* @param groupInfo 群聊信息
|
||
* @param functionInfo 功能信息
|
||
* @param requestId 请求ID
|
||
* @returns
|
||
*/
|
||
const parseGroupAgentQuery = async (
|
||
userInput: string,
|
||
groupInfo: string,
|
||
functionInfo: string,
|
||
requestId: string
|
||
) => {
|
||
const { langfuseHandler } = await getLangfuse(
|
||
"parseGroupAgentQuery",
|
||
requestId
|
||
)
|
||
const model = await getModel()
|
||
const structuredLlm = model.withStructuredOutput(groupAgentConfig, {
|
||
name: "groupAgent",
|
||
})
|
||
return await structuredLlm.invoke(
|
||
`
|
||
当前时间为：${new Date().toLocaleString("zh-CN", { timeZone: "Asia/Shanghai" })}
|
||
所有可用群组信息：${groupInfo}
|
||
所有支持功能信息：${functionInfo}
|
||
你是一个专业的语义解析工程师，给定以下用户输入，帮我解析出群聊ID、群聊名称、功能ID、功能名称、开始时间和结束时间。
|
||
默认功能为总结消息，时间如果用户没有输入则留空
|
||
返回格式定义为：
|
||
\`\`\`
|
||
{
|
||
"chatId": "oc_ef98c2a9229657f99d4ef573a30fe91c",
|
||
"chatName": "MIAI-FE 人工智能部-前端组",
|
||
"functionId": "summary-qwen-72b-instruct-int4",
|
||
"functionName": "总结消息",
|
||
"startTime": "2022-01-01 00:00:00",
|
||
"endTime": "2022-01-01 23:59:59"
|
||
}
|
||
\`\`\`
|
||
如果不包含对应内容，请返回空值。
|
||
|
||
用户输入：
|
||
\`\`\`
|
||
${userInput.replaceAll("`", " ")}
|
||
\`\`\`
|
||
`,
|
||
{
|
||
callbacks: [langfuseHandler],
|
||
}
|
||
)
|
||
}
|
||
|
||
/**
|
||
* 调用LLM模型
|
||
* @param promptName 提示Key
|
||
* @param variables 变量
|
||
* @param requestId 请求ID
|
||
* @param temperature 温度
|
||
* @returns
|
||
*/
|
||
const invoke = async (
|
||
promptName: string,
|
||
variables: Record<string, any>,
|
||
requestId: string,
|
||
temperature = 0
|
||
) => {
|
||
const { langfuse, langfuseHandler } = await getLangfuse("invoke", requestId)
|
||
const prompt = await langfuse.getPrompt(promptName)
|
||
|
||
const langchainTextPrompt = PromptTemplate.fromTemplate(
|
||
prompt.getLangchainPrompt()
|
||
).withConfig({
|
||
metadata: { langfusePrompt: prompt },
|
||
})
|
||
|
||
const chain = langchainTextPrompt.pipe(await getModel(temperature))
|
||
|
||
const { content } = await chain.invoke(variables, {
|
||
callbacks: [langfuseHandler],
|
||
})
|
||
|
||
return content
|
||
}
|
||
|
||
const llm = {
|
||
parseGroupAgentQuery,
|
||
invoke,
|
||
}
|
||
|
||
export default llm
|