vector-db-setup

向量数据库设置

配置向量数据库，用于语义搜索和 AI 应用。

核心工作流

1. 选择数据库：根据需求进行选择
2. 设置连接：配置客户端
3. 生成嵌入：创建向量表示
4. 索引文档：存储文档及其元数据
5. 查询向量：进行语义相似度搜索
6. 优化：调整以提升性能

数据库对比

数据库	类型	最适合	扩展方式
Pinecone	托管	生产环境，无需运维	自动
Chroma	嵌入式/服务端	开发、本地使用	手动
pgvector	PostgreSQL 扩展	已有 Postgres 的环境	随 Postgres 扩展
Qdrant	自托管	完全掌控	手动
Weaviate	托管/自托管	类 GraphQL API	两种方式均可

生成嵌入

OpenAI 嵌入

// embeddings/openai.ts
import OpenAI from 'openai';

const openai = new OpenAI();

export async function generateEmbedding(text: string): Promise<number[]> {
  const response = await openai.embeddings.create({
    model: 'text-embedding-3-small', // or text-embedding-3-large
    input: text,
  });

  return response.data[0].embedding;
}

export async function generateEmbeddings(texts: string[]): Promise<number[][]> {
  const response = await openai.embeddings.create({
    model: 'text-embedding-3-small',
    input: texts,
  });

  return response.data.map((d) => d.embedding);
}

批处理

// embeddings/batch.ts
const BATCH_SIZE = 100;

export async function batchGenerateEmbeddings(
  texts: string[]
): Promise<number[][]> {
  const embeddings: number[][] = [];

  for (let i = 0; i < texts.length; i += BATCH_SIZE) {
    const batch = texts.slice(i, i + BATCH_SIZE);
    const batchEmbeddings = await generateEmbeddings(batch);
    embeddings.push(...batchEmbeddings);

    // Rate limiting
    if (i + BATCH_SIZE < texts.length) {
      await new Promise((resolve) => setTimeout(resolve, 100));
    }
  }

  return embeddings;
}

Pinecone 设置

安装与配置

npm install @pinecone-database/pinecone

// db/pinecone.ts
import { Pinecone } from '@pinecone-database/pinecone';

const pinecone = new Pinecone({
  apiKey: process.env.PINECONE_API_KEY!,
});

// Get or create index
export async function getIndex(indexName: string) {
  const indexes = await pinecone.listIndexes();

  if (!indexes.indexes?.find((i) => i.name === indexName)) {
    await pinecone.createIndex({
      name: indexName,
      dimension: 1536, // OpenAI embedding dimension
      metric: 'cosine',
      spec: {
        serverless: {
          cloud: 'aws',
          region: 'us-east-1',
        },
      },
    });

    // Wait for index to be ready
    await new Promise((resolve) => setTimeout(resolve, 60000));
  }

  return pinecone.Index(indexName);
}

Upsert 与查询

// db/pinecone-ops.ts
import { getIndex } from './pinecone';
import { generateEmbedding, generateEmbeddings } from '../embeddings/openai';

const index = await getIndex('my-index');

interface Document {
  id: string;
  content: string;
  metadata: Record<string, any>;
}

// Upsert documents
export async function upsertDocuments(
  documents: Document[],
  namespace = 'default'
) {
  const embeddings = await generateEmbeddings(documents.map((d) => d.content));

  const vectors = documents.map((doc, i) => ({
    id: doc.id,
    values: embeddings[i],
    metadata: {
      content: doc.content,
      ...doc.metadata,
    },
  }));

  // Upsert in batches
  const BATCH_SIZE = 100;
  for (let i = 0; i < vectors.length; i += BATCH_SIZE) {
    const batch = vectors.slice(i, i + BATCH_SIZE);
    await index.namespace(namespace).upsert(batch);
  }
}

// Query similar documents
export async function querySimilar(
  query: string,
  options: {
    topK?: number;
    namespace?: string;
    filter?: Record<string, any>;
  } = {}
) {
  const { topK = 5, namespace = 'default', filter } = options;

  const queryEmbedding = await generateEmbedding(query);

  const results = await index.namespace(namespace).query({
    vector: queryEmbedding,
    topK,
    includeMetadata: true,
    filter,
  });

  return results.matches?.map((match) => ({
    id: match.id,
    score: match.score,
    content: match.metadata?.content,
    metadata: match.metadata,
  }));
}

// Delete documents
export async function deleteDocuments(ids: string[], namespace = 'default') {
  await index.namespace(namespace).deleteMany(ids);
}

// Delete by filter
export async function deleteByFilter(
  filter: Record<string, any>,
  namespace = 'default'
) {
  await index.namespace(namespace).deleteMany({ filter });
}

Chroma 设置

安装与配置

npm install chromadb

// db/chroma.ts
import { ChromaClient, OpenAIEmbeddingFunction } from 'chromadb';

const client = new ChromaClient({
  path: process.env.CHROMA_URL || 'http://localhost:8000',
});

const embedder = new OpenAIEmbeddingFunction({
  openai_api_key: process.env.OPENAI_API_KEY!,
  openai_model: 'text-embedding-3-small',
});

export async function getCollection(name: string) {
  return client.getOrCreateCollection({
    name,
    embeddingFunction: embedder,
    metadata: { 'hnsw:space': 'cosine' },
  });
}

Chroma 操作

// db/chroma-ops.ts
import { getCollection } from './chroma';

const collection = await getCollection('documents');

// Add documents (Chroma generates embeddings)
export async function addDocuments(documents: Document[]) {
  await collection.add({
    ids: documents.map((d) => d.id),
    documents: documents.map((d) => d.content),
    metadatas: documents.map((d) => d.metadata),
  });
}

// Query
export async function query(queryText: string, nResults = 5) {
  const results = await collection.query({
    queryTexts: [queryText],
    nResults,
  });

  return results.ids[0].map((id, i) => ({
    id,
    content: results.documents?.[0][i],
    metadata: results.metadatas?.[0][i],
    distance: results.distances?.[0][i],
  }));
}

// Query with filter
export async function queryWithFilter(
  queryText: string,
  filter: Record<string, any>,
  nResults = 5
) {
  const results = await collection.query({
    queryTexts: [queryText],
    nResults,
    where: filter,
  });

  return results;
}

// Update document
export async function updateDocument(id: string, content: string, metadata?: Record<string, any>) {
  await collection.update({
    ids: [id],
    documents: [content],
    metadatas: metadata ? [metadata] : undefined,
  });
}

// Delete
export async function deleteDocuments(ids: string[]) {
  await collection.delete({ ids });
}

pgvector 设置

安装

npm install pg pgvector

-- Enable extension
CREATE EXTENSION vector;

-- Create table
CREATE TABLE documents (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  content TEXT NOT NULL,
  metadata JSONB,
  embedding vector(1536),
  created_at TIMESTAMP DEFAULT NOW()
);

-- Create index for similarity search
CREATE INDEX ON documents USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 100);

-- Or use HNSW (better for production)
CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops);

pgvector 操作

// db/pgvector.ts
import { Pool } from 'pg';
import pgvector from 'pgvector/pg';

const pool = new Pool({
  connectionString: process.env.DATABASE_URL,
});

// Register pgvector type
await pgvector.registerType(pool);

// Insert document
export async function insertDocument(
  content: string,
  embedding: number[],
  metadata?: Record<string, any>
) {
  const result = await pool.query(
    `INSERT INTO documents (content, embedding, metadata)
     VALUES ($1, $2, $3)
     RETURNING id`,
    [content, pgvector.toSql(embedding), metadata]
  );

  return result.rows[0].id;
}

// Similarity search
export async function searchSimilar(
  queryEmbedding: number[],
  limit = 5,
  threshold = 0.7
) {
  const result = await pool.query(
    `SELECT id, content, metadata,
            1 - (embedding <=> $1) as similarity
     FROM documents
     WHERE 1 - (embedding <=> $1) > $2
     ORDER BY embedding <=> $1
     LIMIT $3`,
    [pgvector.toSql(queryEmbedding), threshold, limit]
  );

  return result.rows;
}

// Search with metadata filter
export async function searchWithFilter(
  queryEmbedding: number[],
  filter: Record<string, any>,
  limit = 5
) {
  const result = await pool.query(
    `SELECT id, content, metadata,
            1 - (embedding <=> $1) as similarity
     FROM documents
     WHERE metadata @> $2
     ORDER BY embedding <=> $1
     LIMIT $3`,
    [pgvector.toSql(queryEmbedding), filter, limit]
  );

  return result.rows;
}

// Hybrid search (vector + full-text)
export async function hybridSearch(
  queryEmbedding: number[],
  textQuery: string,
  limit = 5
) {
  const result = await pool.query(
    `SELECT id, content, metadata,
            (1 - (embedding <=> $1)) * 0.7 +
            ts_rank(to_tsvector(content), plainto_tsquery($2)) * 0.3 as score
     FROM documents
     WHERE to_tsvector(content) @@ plainto_tsquery($2)
        OR 1 - (embedding <=> $1) > 0.5
     ORDER BY score DESC
     LIMIT $3`,
    [pgvector.toSql(queryEmbedding), textQuery, limit]
  );

  return result.rows;
}

Qdrant 设置

npm install @qdrant/js-client-rest

// db/qdrant.ts
import { QdrantClient } from '@qdrant/js-client-rest';

const client = new QdrantClient({
  url: process.env.QDRANT_URL,
  apiKey: process.env.QDRANT_API_KEY,
});

// Create collection
export async function createCollection(name: string) {
  await client.createCollection(name, {
    vectors: {
      size: 1536,
      distance: 'Cosine',
    },
  });
}

// Upsert points
export async function upsertPoints(
  collection: string,
  points: Array<{
    id: string;
    vector: number[];
    payload: Record<string, any>;
  }>
) {
  await client.upsert(collection, {
    points: points.map((p) => ({
      id: p.id,
      vector: p.vector,
      payload: p.payload,
    })),
  });
}

// Search
export async function search(
  collection: string,
  vector: number[],
  limit = 5,
  filter?: Record<string, any>
) {
  const results = await client.search(collection, {
    vector,
    limit,
    filter: filter
      ? {
          must: Object.entries(filter).map(([key, value]) => ({
            key,
            match: { value },
          })),
        }
      : undefined,
    with_payload: true,
  });

  return results;
}

文档处理流水线

// pipeline/ingest.ts
import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';
import { generateEmbeddings } from '../embeddings/openai';
import { upsertDocuments } from '../db/pinecone-ops';

interface RawDocument {
  id: string;
  content: string;
  source: string;
  metadata?: Record<string, any>;
}

export async function ingestDocuments(documents: RawDocument[]) {
  const splitter = new RecursiveCharacterTextSplitter({
    chunkSize: 1000,
    chunkOverlap: 200,
  });

  const chunks: Array<{
    id: string;
    content: string;
    metadata: Record<string, any>;
  }> = [];

  for (const doc of documents) {
    const splits = await splitter.splitText(doc.content);

    splits.forEach((text, index) => {
      chunks.push({
        id: `${doc.id}-chunk-${index}`,
        content: text,
        metadata: {
          source: doc.source,
          documentId: doc.id,
          chunkIndex: index,
          ...doc.metadata,
        },
      });
    });
  }

  // Upsert in batches
  await upsertDocuments(chunks);

  return { totalChunks: chunks.length };
}

最佳实践

1. 选择合适的维度：与嵌入模型相匹配
2. 使用命名空间：以符合逻辑的方式组织数据
3. 添加元数据：支持筛选
4. 批量操作：减少 API 调用
5. 合理建立索引：使用 HNSW 提升速度，使用 IVF 节省内存
6. 监控性能：跟踪延迟和召回率
7. 缓存嵌入：避免重复生成
8. 使用混合搜索：结合向量搜索和关键词搜索

输出检查清单

每个向量数据库设置都应包含：

• 已配置数据库客户端
• 嵌入生成函数
• 集合/索引创建
• 带元数据的文档 upsert
• 相似度搜索函数
• 筛选搜索功能
• 面向大型数据集的批处理
• 删除/更新操作
• 错误处理
• 性能监控