RAG & 검색 증강 생성

import { embedMany } from 'ai'
import { handleCallback } from '@vercel/queue'

export const POST = handleCallback(async (job) => {
  const { chunks, docId, metadata } = job

  const { embeddings } = await embedMany({
    model: 'openai/text-embedding-3-small',
    values: chunks,
    providerOptions: {
      gateway: { tags: ['feature:rag-indexing', `doc:${docId}`] },
    },
  })

  await vectorStore.upsert(
    chunks.map((chunk, i) => ({
      id: `${docId}#${i}`,
      values: embeddings[i],
      metadata: { ...metadata, chunk },
    }))
  )
})

import { generateText, tool } from 'ai'
import { z } from 'zod'

const searchDocs = tool({
  description: '사내 문서를 검색합니다.',
  inputSchema: z.object({ query: z.string() }),
  execute: async ({ query }) => {
    // 1. Vector search: top-k 후보
    const candidates = await vectorStore.search({
      query: await embed(query),
      topK: 20,
    })

    // 2. Rerank: 정밀도 향상
    const reranked = await reranker.rank({
      query,
      documents: candidates.map((c) => c.metadata.chunk),
      topK: 5,
    })

    return reranked.map((r) => ({
      content: r.document,
      source: candidates[r.index].metadata.source,
      score: r.score,
    }))
  },
})

- docs-server (MCP): resources로 검색 결과 공급
  - searchDocs: 사내 문서 검색
  - getDocument: 특정 문서 조회
  - 모두 read-only, 앱이 주도