1
export type Bindings = {
2
  DOCS_INDEX: Vectorize  // 和 Ai 一样，Cloudflare Workers 内置的全局类型
3
  AI: Ai
4
}
5

6
export type AppEnv = {
7
  Bindings: Bindings
8
}

01
import { Hono } from 'hono'
02
import type { AppEnv } from '../types'
03

04
const ingest = new Hono<AppEnv>()
05

06
ingest.post('/ingest', async (c) => {
07
  const { docs } = await c.req.json<{
08
    docs: Array<{ id: string; text: string; source?: string }>
09
  }>()
10

11
  // 1. 批量生成 embedding（Workers AI 一次最多 100 段文本）
12
  //    返回 { data: [向量1, 向量2, ...] }，每个向量是 float[]
13
  const texts = docs.map((d) => d.text)
14
  const { data: embeddings } = await c.env.AI.run('@cf/baai/bge-base-en-v1.5', {
15
    text: texts,
16
  })
17

18
  // 2. 组装成 Vectorize 的格式
19
  const vectors = docs.map((doc, i) => ({
20
    id: doc.id,
21
    values: embeddings[i],
22
    metadata: {
23
      text: doc.text,
24
      source: doc.source || 'unknown',
25
      insertedAt: Date.now(),
26
    },
27
  }))
28

29
  // 3. 一次批量写入
30
  const result = await c.env.DOCS_INDEX.upsert(vectors)
31

32
  return c.json({
33
    ingested: vectors.length,
34
    mutationId: result.mutationId,
35
  })
36
})
37

38
export default ingest

01
import { Hono } from 'hono'
02
import type { AppEnv } from '../types'
03

04
const search = new Hono<AppEnv>()
05

06
search.post('/search', async (c) => {
07
  const { query, topK = 5 } = await c.req.json<{
08
    query: string
09
    topK?: number
10
  }>()
11

12
  // 1. 把问题也转成向量
13
  const { data: queryEmbeddings } = await c.env.AI.run(
14
    '@cf/baai/bge-base-en-v1.5',
15
    { text: [query] }
16
  )
17
  const queryVector = queryEmbeddings[0]
18

19
  // 2. 查最相似的 topK 条
20
  const result = await c.env.DOCS_INDEX.query(queryVector, {
21
    topK,
22
    returnMetadata: 'all',
23
  })
24

25
  return c.json({
26
    matches: result.matches.map((m) => ({
27
      id: m.id,
28
      score: m.score,  // 相似度分数，0~1 之间
29
      text: m.metadata?.text,
30
      source: m.metadata?.source,
31
    })),
32
  })
33
})
34

35
export default search

01
import { Hono } from 'hono'
02
import { streamSSE } from 'hono/streaming'
03
import type { AppEnv } from '../types'
04

05
const rag = new Hono<AppEnv>()
06

07
rag.post('/rag', async (c) => {
08
  const { question } = await c.req.json<{ question: string }>()
09

10
  // 1. 检索相关文档
11
  //    只传了一段文本，所以 data 里只有一个向量，直接解构出来
12
  const { data: [queryVec] } = await c.env.AI.run(
13
    '@cf/baai/bge-base-en-v1.5',
14
    { text: [question] }
15
  )
16

17
  const { matches } = await c.env.DOCS_INDEX.query(queryVec, {
18
    topK: 3,
19
    returnMetadata: 'all',
20
  })
21

22
  // 2. 把检索到的片段拼进 prompt
23
  const context = matches
24
    .map((m, i) => `[${i + 1}] ${m.metadata?.text}`)
25
    .join('\n\n')
26

27
  const systemPrompt = `你是一个根据提供的上下文回答问题的助手。
28
只使用下面的上下文回答。如果上下文里没有答案，就回复"根据现有资料无法回答"。
29

30
上下文：
31
${context}`
32

33
  // 3. 调用大模型流式生成
34
  return streamSSE(c, async (stream) => {
35
    const llmStream = await c.env.AI.run(
36
      '@cf/meta/llama-3.1-8b-instruct',
37
      {
38
        messages: [
39
          { role: 'system', content: systemPrompt },
40
          { role: 'user', content: question },
41
        ],
42
        stream: true,
43
      }
44
    )
45

46
    const reader = (llmStream as ReadableStream).getReader()
47
    const decoder = new TextDecoder()
48

49
    while (true) {
50
      const { done, value } = await reader.read()
51
      if (done) break
52
      await stream.writeSSE({
53
        data: decoder.decode(value, { stream: true }),
54
        event: 'delta',
55
      })
56
    }
57

58
    // 顺便把引用来源也发给前端，方便展示
59
    await stream.writeSSE({
60
      data: JSON.stringify(matches.map((m) => ({ id: m.id, score: m.score }))),
61
      event: 'sources',
62
    })
63
  })
64
})
65

66
export default rag

1
const result = await c.env.DOCS_INDEX.query(queryVector, {
2
  topK: 5,
3
  returnMetadata: 'all',
4
  filter: {
5
    source: { $eq: 'internal-wiki' },
6
  },
7
})

1
const result = await c.env.DOCS_INDEX.query(queryVector, {
2
  topK: 5,
3
  returnMetadata: 'all',
4
  filter: {
5
    tenantId: { $eq: c.get('user').tenantId },
6
  },
7
})

1
// 写入时
2
await c.env.DOCS_INDEX.upsert(vectors.map(v => ({ ...v, namespace: 'tenant-42' })))
3

4
// 查询时
5
const result = await c.env.DOCS_INDEX.query(queryVec, {
6
  namespace: 'tenant-42',
7
  topK: 5,
8
})