1# Workers AI Patterns
2 
3## RAG (Retrieval-Augmented Generation)
4 
5```typescript
6// 1. Embed query
7const embedding = await env.AI.run('@cf/baai/bge-base-en-v1.5', { text: query });
8 
9// 2. Search vectors
10const results = await env.VECTORIZE.query(embedding.data[0], {
11  topK: 5, returnMetadata: true
12});
13 
14// 3. Build context
15const context = results.matches.map(m => m.metadata?.text).join('\n\n');
16 
17// 4. Generate with context
18const response = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
19  messages: [
20    { role: 'system', content: `Answer based on:\n\n${context}` },
21    { role: 'user', content: query }
22  ]
23});
24```
25 
26## Streaming (SSE)
27 
28```typescript
29const stream = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
30  messages, stream: true
31});
32 
33const { readable, writable } = new TransformStream();
34const writer = writable.getWriter();
35 
36(async () => {
37  for await (const chunk of stream) {
38    await writer.write(new TextEncoder().encode(`data: ${JSON.stringify(chunk)}\n\n`));
39  }
40  await writer.write(new TextEncoder().encode('data: [DONE]\n\n'));
41  await writer.close();
42})();
43 
44return new Response(readable, {
45  headers: { 'Content-Type': 'text/event-stream' }
46});
47```
48 
49## Error Handling & Retry
50 
51```typescript
52async function runWithRetry(env, model, input, maxRetries = 3) {
53  for (let attempt = 0; attempt < maxRetries; attempt++) {
54    try {
55      return await env.AI.run(model, input);
56    } catch (error) {
57      if (error.message?.includes('7505') && attempt < maxRetries - 1) {
58        await new Promise(r => setTimeout(r, Math.pow(2, attempt) * 1000));
59        continue;
60      }
61      throw error;
62    }
63  }
64}
65```
66 
67## Model Fallback
68 
69```typescript
70try {
71  return await env.AI.run('@cf/meta/llama-3.1-70b-instruct', { messages });
72} catch {
73  return await env.AI.run('@cf/meta/llama-3.1-8b-instruct', { messages });
74}
75```
76 
77## Prompt Patterns
78 
79```typescript
80// System prompts
81const PROMPTS = {
82  json: 'Respond with valid JSON only.',
83  concise: 'Keep responses brief.',
84  cot: 'Think step by step before answering.'
85};
86 
87// Few-shot
88messages: [
89  { role: 'system', content: 'Extract as JSON' },
90  { role: 'user', content: 'John bought 3 apples for $5' },
91  { role: 'assistant', content: '{"name":"John","item":"apples","qty":3}' },
92  { role: 'user', content: actualInput }
93]
94```
95 
96## Parallel Execution
97 
98```typescript
99const [sentiment, summary, embedding] = await Promise.all([
100  env.AI.run('@cf/mistral/mistral-7b-instruct-v0.1', { messages: sentimentPrompt }),
101  env.AI.run('@cf/meta/llama-3.1-8b-instruct', { messages: summaryPrompt }),
102  env.AI.run('@cf/baai/bge-base-en-v1.5', { text })
103]);
104```
105 
106## Cost Optimization
107 
108| Task | Model | Neurons |
109|------|-------|---------|
110| Classify | `@cf/mistral/mistral-7b-instruct-v0.1` | ~50 |
111| Chat | `@cf/meta/llama-3.1-8b-instruct` | ~200 |
112| Complex | `@cf/meta/llama-3.1-70b-instruct` | ~2000 |
113| Embed | `@cf/baai/bge-base-en-v1.5` | ~10 |
114 
115```typescript
116// Batch embeddings
117const response = await env.AI.run('@cf/baai/bge-base-en-v1.5', {
118  text: textsArray // Process multiple at once
119});
120```
121