从 RAG to Riches：用 Meilisearch 混合搜索构建更可靠的检索增强生成

npm install meilisearch
# 或
yarn add meilisearch

import { Meilisearch } from 'meilisearch'

const client = new Meilisearch({
  host: process.env.MEILISEARCH_URL,  // 例如 http://localhost:7700
  apiKey: process.env.MEILISEARCH_KEY
})

const index = client.index('movies')

// 配置 embedder，指定使用 OpenAI 的 embedding 模型
await index.updateEmbedders({
  movies: {
    source: 'openAi',
    apiKey: process.env.OPENAI_API_KEY,
    model: 'text-embedding-3-small',
    // 使用 Liquid 模板聚焦最有检索价值的字段
    documentTemplate: "{{doc.title}}：{{doc.overview}}。类型：{{doc.genre}}"
  }
})

const movies = [
  { id: 1, title: '盗梦空间', overview: '一个关于梦境窃取和植入的科幻惊悚片', genre: '科幻' },
  { id: 2, title: '星际穿越', overview: '宇航员穿越虫洞寻找人类新家园', genre: '科幻' },
  { id: 3, title: '当幸福来敲门', overview: '一位父亲在困境中追求幸福的励志故事', genre: '剧情' },
]

const task = await index.addDocuments(movies)
await client.waitForTask(task.taskUid)

console.log('文档已添加并完成嵌入生成')

// 混合搜索：semanticRatio=0.5 表示关键词和语义各占一半
const results = await index.search('适合全家一起看的科幻电影', {
  hybrid: {
    embedder: 'movies',
    semanticRatio: 0.5  // 0.0 = 纯关键词, 1.0 = 纯语义
  },
  limit: 5
})

for (const hit of results.hits) {
  console.log(`${hit.title} (Score: ${hit._rankingScore?.toFixed(4) ?? 'N/A'})`)
}

import { Meilisearch } from 'meilisearch'
import { OpenAIEmbeddings } from '@langchain/openai'
import { MeilisearchVectorStore } from '@langchain/community/vectorstores/meilisearch'

const embeddings = new OpenAIEmbeddings()
const client = new Meilisearch({
  host: process.env.MEILISEARCH_URL,
  apiKey: process.env.MEILISEARCH_KEY
})

// 创建 vector store 并索引文档
const vectorStore = await MeilisearchVectorStore.fromTexts(
  ['盗梦空间是一部关于梦境窃取的科幻片', '星际穿越讲述宇航员穿越虫洞的故事'],
  [{ genre: '科幻' }, { genre: '科幻' }],
  embeddings,
  {
    client,
    indexName: 'movies_rag'
  }
)

// 相似度搜索
const docs = await vectorStore.similaritySearch('适合全家看的科幻电影', 3)
docs.forEach(doc => console.log(doc.pageContent))

const query = 'wireless ergonomic keyboard'

const ratios = [
  { ratio: 0.0, desc: '纯关键词搜索' },
  { ratio: 0.5, desc: '均衡混合' },
  { ratio: 1.0, desc: '纯语义搜索' },
]

for (const { ratio, desc } of ratios) {
  console.log(`\n--- ${desc} (semanticRatio: ${ratio}) ---`)
  
  const results = await index.search(query, {
    hybrid: {
      embedder: 'products',
      semanticRatio: ratio
    },
    limit: 5
  })
  
  results.hits.forEach((hit, i) => {
    console.log(`${i + 1}. ${hit.title} (${hit._rankingScore?.toFixed(4)})`)
  })
}