upstash-vector-db-skills（Upstash 矢量数据库技能）

Links

• Docs: https://upstash.com/docs/vector
• Getting Started: https://upstash.com/docs/vector/overall/getstarted
• Semantic Search Tutorial: https://upstash.com/docs/vector/tutorials/semantic_search
• Namespaces: https://upstash.com/docs/vector/features/namespaces
• Embedding Models: https://upstash.com/docs/vector/features/embeddingmodels
• MixBread AI: https://www.mixbread.ai/ (preferred embedding provider)

Quick Setup

1. Create Vector Index (Upstash Console)

• Go to Upstash Console
• Create Vector Index: name, region (closest to app), type (Dense for semantic search)
• Select embedding model: MixBread AI recommended (or use Upstash built-in models)
• Copy UPSTASH_VECTOR_REST_URL and UPSTASH_VECTOR_REST_TOKEN to .env

2. Install SDK

pnpm add @upstash/vector

3. Environment

UPSTASH_VECTOR_REST_URL=your_url
UPSTASH_VECTOR_REST_TOKEN=your_token

Code Examples

Initialize Client (Node.js / TypeScript)

import { Index } from "@upstash/vector";

const index = new Index({
  url: process.env.UPSTASH_VECTOR_REST_URL,
  token: process.env.UPSTASH_VECTOR_REST_TOKEN,
});

Upsert Documents (Auto-Embed)

When using an embedding model in the index, text is embedded automatically:

// Single document
await index.upsert({
  id: "doc-1",
  data: "Upstash provides serverless vector database solutions.",
  metadata: { source: "docs", category: "intro" },
});

// Batch
await index.upsert([
  { id: "doc-2", data: "Vector search powers semantic similarity.", metadata: { source: "docs" } },
  { id: "doc-3", data: "MixBread AI provides high-quality embeddings.", metadata: { source: "blog" } },
]);

Query / Semantic Search

// Semantic search with auto-embedding
const results = await index.query({
  data: "What is semantic search?",
  topK: 5,
  includeMetadata: true,
});

results.forEach((result) => {
  console.log(`ID: ${result.id}, Score: ${result.score}, Metadata:`, result.metadata);
});

Using Namespaces (Data Isolation)

Namespaces partition a single index into isolated subsets. Useful for multi-tenant or multi-domain apps.

// Upsert in namespace "blog"
await index.namespace("blog").upsert({
  id: "post-1",
  data: "Next.js tutorial for Vercel deployment",
  metadata: { author: "user-123" },
});

// Query only "blog" namespace
const blogResults = await index.namespace("blog").query({
  data: "Vercel deployment",
  topK: 3,
  includeMetadata: true,
});

// List all namespaces
const namespaces = await index.listNamespaces();
console.log(namespaces);

// Delete namespace
await index.deleteNamespace("blog");

Full Semantic Search Example (Vercel Function)

// api/search.ts (Vercel Edge Function or Serverless Function)
import { Index } from "@upstash/vector";

export const config = {
  runtime: "nodejs", // or "edge"
};

const index = new Index({
  url: process.env.UPSTASH_VECTOR_REST_URL,
  token: process.env.UPSTASH_VECTOR_REST_TOKEN,
});

export default async function handler(req, res) {
  if (req.method !== "POST") {
    return res.status(405).json({ error: "Method not allowed" });
  }

  const { query, namespace = "", topK = 5 } = req.body;

  try {
    const searchIndex = namespace ? index.namespace(namespace) : index;
    const results = await searchIndex.query({
      data: query,
      topK,
      includeMetadata: true,
    });

    return res.status(200).json({ results });
  } catch (error) {
    console.error("Search error:", error);
    return res.status(500).json({ error: "Search failed" });
  }
}

Index Operations

// Reset (clear all vectors in index or namespace)
await index.reset();

// Or reset a specific namespace
await index.namespace("old-data").reset();

// Delete a single vector
await index.delete("doc-1");

// Delete multiple vectors
await index.delete(["doc-1", "doc-2", "doc-3"]);

Embedding Models

Available in Upstash

• BAAI/bge-large-en-v1.5 (1024 dim, best performance, ~64.23 MTEB score)
• BAAI/bge-base-en-v1.5 (768 dim, good balance)
• BAAI/bge-small-en-v1.5 (384 dim, lightweight)
• BAAI/bge-m3 (1024 dim, sparse + dense hybrid)

Recommended: MixBread AI

If using MixBread as your embedding provider:

1. Create a MixBread API key at https://www.mixbread.ai/
2. When creating your Upstash index, select MixBread as the embedding model.
3. MixBread handles tokenization and semantic quality automatically.
4. No extra setup needed in your code; use index.upsert() / index.query() with text directly.

Best Practices

For Vercel Deployment

• Store credentials in Vercel Environment Variables (project settings or .env.local).
• Use Edge Functions or Serverless Functions for low-latency access.
• Implement request rate limiting to stay within Upstash quotas.

Namespace Strategy

• Use namespaces to isolate data by tenant, domain, or use case.
• Example: namespace("user-123") for per-user search.
• Clean up old namespaces to avoid storage bloat.

Query Performance

• Keep topK reasonable (5–10 typically sufficient).
• Use metadata filtering to pre-filter results if possible.
• Upstash is eventually consistent; expect slight delays after upserts.

Error Handling

try {
  const results = await index.query({
    data: userQuery,
    topK: 5,
    includeMetadata: true,
  });
} catch (error) {
  if (error.status === 401) {
    console.error("Invalid credentials");
  } else if (error.status === 429) {
    console.error("Rate limited");
  } else {
    console.error("Query error:", error);
  }
}

Common Patterns

RAG (Retrieval Augmented Generation)

1. Upsert documents / knowledge base into Upstash.
2. On user query, retrieve top-k similar docs via semantic search.
3. Pass retrieved docs + user query to LLM for better context.

const docs = await index.query({ data: userQuestion, topK: 3 });
const context = docs.map((d) => d.metadata?.text).join("\n");
// Pass context to LLM

Multi-Tenant Search

Use namespaces to isolate each tenant's vectors:

const userNamespace = `tenant-${userId}`;
await index.namespace(userNamespace).upsert({ id, data, metadata });
// Queries only see that tenant's data

Batch Indexing

For bulk imports, upsert in batches:

const batchSize = 100;
for (let i = 0; i < documents.length; i += batchSize) {
  const batch = documents.slice(i, i + batchSize);
  await index.upsert(batch);
  console.log(`Indexed batch ${i / batchSize + 1}`);
}

Troubleshooting

• No results returned: Ensure documents are indexed and embedding model is active.
• Slow queries: Check quota limits; consider upgrading plan or reducing dataset size.
• Stale data: Upstash is eventually consistent; wait 1–2 seconds before querying new inserts.
• Namespace not working: Ensure namespace exists (created on first upsert) or use the default "".