查询向量

Perform similarity search and retrieve vectors using JavaScript SDK or Postgres.

此功能处于测试阶段

预计会有快速变化、功能有限，并可能出现破坏性更新。随着我们改进体验并扩大访问，欢迎分享反馈。

🌐 Expect rapid changes, limited features, and possible breaking updates. Share feedback as we refine the experience and expand access.

向量相似度搜索使用距离度量找到与查询向量最相似的向量。你可以用 JavaScript SDK 查询向量，也可以直接用 SQL 从 Postgres 查询。

🌐 Vector similarity search finds vectors most similar to a query vector using distance metrics. You can query vectors using the JavaScript SDK or directly from Postgres using SQL.

与 pgvector 的比较

向量桶和它们使用的任何外部数据封装器 (FDW) 只支持一种相似性搜索算法，即 <===> 距离运算符。

🌐 Vector buckets and any Foreign Data Wrappers (FDW) they use only support one similarity search algorithm, the <===> distance operator.

基础相似性搜索 #

🌐 Basic similarity search

1
import { createClient } from '@supabase/supabase-js'
2
3
const supabase = createClient('https://your-project-id.supabase.co', 'your-service-key')
4
5
const index = supabase.storage.vectors.from('embeddings').index('documents-openai')
6
7
// Query with a vector embedding
8
const { data, error } = await index.queryVectors({
9
  queryVector: {
10
    float32: [0.1, 0.2, 0.3 /* ... embedding of 1536 dimensions ... */],
11
  },
12
  topK: 5,
13
  returnDistance: true,
14
  returnMetadata: true,
15
})
16
17
if (error) {
18
  console.error('Query failed:', error)
19
} else {
20
  // Results are ranked by similarity (lowest distance = most similar)
21
  data.vectors.forEach((result, rank) => {
22
    console.log(`${rank + 1}. ${result.metadata?.title}`)
23
    console.log(`   Similarity score: ${result.distance.toFixed(4)}`)
24
  })
25
}

语义搜索 #

🌐 Semantic search

通过将查询文本嵌入来查找与查询相似的文档：

🌐 Find documents similar to a query by embedding the query text:

1
import { createClient } from '@supabase/supabase-js'
2
import OpenAI from 'openai'
3
4
const supabase = createClient(...)
5
const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY })
6
7
async function semanticSearch(query, topK = 5) {
8
  // Embed the query
9
  const queryEmbedding = await openai.embeddings.create({
10
    model: 'text-embedding-3-small',
11
    input: query
12
  })
13
14
  const queryVector = queryEmbedding.data[0].embedding
15
16
  // Search for similar vectors
17
  const { data, error } = await supabase.storage.vectors
18
    .from('embeddings')
19
    .index('documents-openai')
20
    .queryVectors({
21
      queryVector: { float32: queryVector },
22
      topK,
23
      returnDistance: true,
24
      returnMetadata: true
25
    })
26
27
  if (error) {
28
    throw error
29
  }
30
31
  return data.vectors.map((result) => ({
32
    id: result.key,
33
    title: result.metadata?.title,
34
    similarity: 1 - result.distance, // Convert distance to similarity (0-1)
35
    metadata: result.metadata
36
  }))
37
}
38
39
// Usage
40
const results = await semanticSearch('How do I use vector search?')
41
results.forEach((result) => {
42
  console.log(`${result.title} (${(result.similarity * 100).toFixed(1)}% similar)`)
43
})

筛选相似搜索 #

🌐 Filtered similarity search

1
const index = supabase.storage.vectors
2
  .from('embeddings')
3
  .index('documents-openai')
4
5
// Search with metadata filter
6
const { data } = await index.queryVectors({
7
  queryVector: { float32: [...embedding...] },
8
  topK: 10,
9
  filter: {
10
    // Filter by metadata fields
11
    category: 'electronics',
12
    in_stock: true,
13
    price: { $lte: 500 } // Less than or equal to 500
14
  },
15
  returnDistance: true,
16
  returnMetadata: true
17
})

获取特定向量 #

🌐 Retrieving specific vectors

1
const index = supabase.storage.vectors.from('embeddings').index('documents-openai')
2
3
const { data, error } = await index.getVectors({
4
  keys: ['doc-1', 'doc-2', 'doc-3'],
5
  returnData: true,
6
  returnMetadata: true,
7
})
8
9
if (!error) {
10
  data.vectors.forEach((vector) => {
11
    console.log(`${vector.key}: ${vector.metadata?.title}`)
12
  })
13
}

列出向量 #

🌐 Listing vectors

1
const index = supabase.storage.vectors.from('embeddings').index('documents-openai')
2
3
let nextToken = undefined
4
let pageCount = 0
5
6
do {
7
  const { data, error } = await index.listVectors({
8
    maxResults: 100,
9
    nextToken,
10
    returnData: false, // Don't return embeddings for faster response
11
    returnMetadata: true,
12
  })
13
14
  if (error) break
15
16
  pageCount++
17
  console.log(`Page ${pageCount}: ${data.vectors.length} vectors`)
18
19
  data.vectors.forEach((vector) => {
20
    console.log(`  - ${vector.key}: ${vector.metadata?.title}`)
21
  })
22
23
  nextToken = data.nextToken
24
} while (nextToken)

混合搜索：向量 + 关系数据 #

🌐 Hybrid search: Vectors + relational data

将相似性搜索与 SQL 过滤和连接结合起来：

🌐 Combine similarity search with SQL filtering and joins:

1
async function hybridSearch(queryVector, filters) {
2
  const index = supabase.storage.vectors.from('embeddings').index('documents-openai')
3
4
  // Get similar vectors with filters
5
  const { data: vectorResults } = await index.queryVectors({
6
    queryVector: { float32: queryVector },
7
    topK: 100,
8
    filter: filters,
9
    returnDistance: true,
10
    returnMetadata: true,
11
  })
12
13
  // Get additional details from relational database
14
  const { data: details } = await supabase
15
    .from('documents')
16
    .select('*')
17
    .in(
18
      'id',
19
      vectorResults.vectors.map((v) => v.metadata?.doc_id)
20
    )
21
22
  // Merge results
23
  return vectorResults.vectors.map((vector) => {
24
    const detail = details?.find((d) => d.id === vector.metadata?.doc_id)
25
    return {
26
      ...vector,
27
      ...detail,
28
    }
29
  })
30
}

真实案例 #

🌐 Real-world examples

RAG（检索增强生成） #

🌐 RAG (retrieval-augmented generation)

1
import OpenAI from 'openai'
2
import { createClient } from '@supabase/supabase-js'
3
4
async function retrieveContextForLLM(userQuery) {
5
  const supabase = createClient(...)
6
  const openai = new OpenAI()
7
8
  // 1. Embed the user query
9
  const queryEmbedding = await openai.embeddings.create({
10
    model: 'text-embedding-3-small',
11
    input: userQuery
12
  })
13
14
  // 2. Retrieve relevant documents
15
  const { data: vectorResults } = await supabase.storage.vectors
16
    .from('embeddings')
17
    .index('documents-openai')
18
    .queryVectors({
19
      queryVector: { float32: queryEmbedding.data[0].embedding },
20
      topK: 5,
21
      returnMetadata: true
22
    })
23
24
  // 3. Use vectors to augment LLM prompt
25
  const context = vectorResults.vectors
26
    .map(v => v.metadata?.content || '')
27
    .join('\n\n')
28
29
  const response = await openai.chat.completions.create({
30
    model: 'gpt-4',
31
    messages: [
32
      {
33
        role: 'system',
34
        content: `Use the following context to answer the user's question:\n\n${context}`
35
      },
36
      {
37
        role: 'user',
38
        content: userQuery
39
      }
40
    ]
41
  })
42
43
  return response.choices[0].message.content
44
}

产品推荐 #

🌐 Product recommendations

1
async function recommendProducts(userEmbedding, topK = 5) {
2
  const supabase = createClient(...)
3
4
  // Find similar products
5
  const { data } = await supabase.storage.vectors
6
    .from('embeddings')
7
    .index('products-openai')
8
    .queryVectors({
9
      queryVector: { float32: userEmbedding },
10
      topK,
11
      filter: {
12
        in_stock: true
13
      },
14
      returnMetadata: true
15
    })
16
17
  return data.vectors.map((result) => ({
18
    id: result.metadata?.product_id,
19
    name: result.metadata?.name,
20
    price: result.metadata?.price,
21
    similarity: 1 - result.distance
22
  }))
23
}

在相似性搜索前进行过滤 #

🌐 Filtering before similarity search

1
// Use metadata filters to reduce search scope
2
const { data } = await index.queryVectors({
3
  queryVector,
4
  topK: 100,
5
  filter: {
6
    category: 'electronics', // Pre-filter by category
7
  },
8
})

下一步 #

🌐 Next steps