chatgpt-api/examples/expert-question-answering.ts

import { OpenAIClient } from '@agentic/openai-fetch'
import 'dotenv/config'
import { z } from 'zod'

import { Agentic, SerpAPITool, withHumanFeedback } from '@/index'

async function main() {
  const openai = new OpenAIClient({ apiKey: process.env.OPENAI_API_KEY! })
  const agentic = new Agentic({ openai })

  const question = 'How do I build a product that people will love?'

  const task = withHumanFeedback(
    agentic
      .gpt4(
        `Generate a list of {n} prominent experts that can answer the following question: {{question}}.`
      )
      .tools([new SerpAPITool()])
      .output(
        z.array(
          z.object({
            name: z.string(),
            bio: z.string()
          })
        )
      )
      .input(
        z.object({
          question: z.string(),
          n: z.number().int().default(5)
        })
      ),
    {
      type: 'selectN'
    }
  )
  const { metadata } = await task.callWithMetadata({
    question
  })

  if (
    metadata.feedback &&
    metadata.feedback.type === 'selectN' &&
    metadata.feedback.selected
  ) {
    const answer = agentic
      .gpt4(
        `Generate an answer to the following question: "{{question}}" from each of the following experts: {{#each experts}}
      - {{this.name}}: {{this.bio}}
      {{/each}}`
      )
      .output(
        z.array(
          z.object({
            expert: z.string(),
            answer: z.string()
          })
        )
      )
      .input(
        z.object({
          question: z.string(),
          experts: z.array(z.object({ name: z.string(), bio: z.string() }))
        })
      )
      .call({
        question,
        experts: metadata.feedback.selected
      })
    console.log(answer)
  }
}

main()