ease the integration of hermes function calling prompt format into typescript ecosystem with typed function calls
- generate system message prompt with your defined tools, zod to json schema
- parse tool call output -> typed
pnpm add functhem
import { z } from "zod"
import { Tool, Functhem } from "functhem"
const happySchema = z.object({
name: z.literal("isHappy"),
arguments: z.object({
isHappy: z.boolean(),
}),
})
const happyTool = new Tool<typeof happySchema>({
name: "isHappy",
description: "Is the user happy?",
schema: happySchema,
})
const movieSchema = z.object({
name: z.literal("watchMovie"),
arguments: z.object({
genre: z.string(),
}),
})
const movieTool = new Tool<typeof movieSchema>({
name: "watchMovie",
description:
"Does the user express that they want to see a movie, if so, which genre?",
schema: movieSchema,
})
const functhem = new Functhem([happyTool, movieTool])
const sysMsg = functhem.getSysMsg()
const output = await yourInferenceFunction({
messages: [
{
role: "system",
content: functhem.getSysMsg(), // <-- HERE
},
{
role: "user",
content: `Input: so stressed, just want to see a funny movie\nProcess the user input with your tools.`,
},
],
})
const outCalls = functhem.parseOut(output) // <-- HERE
outCalls.forEach((one) => {
switch (one.name) {
case "isHappy": {
console.log(one.arguments.isHappy)
break
}
case "watchMovie": {
console.log(one.arguments.genre)
break
}
}
})