Prompt Caching deep — kỹ thuật giảm 90% phí AI mà ít dev biết tận dụng

const message = await client.messages.create({
  model: "claude-sonnet-4-5",
  system: [
    {
      type: "text",
      text: longSystemPrompt,  // 5000 token boilerplate
      cache_control: { type: "ephemeral" }
    }
  ],
  messages: [{ role: "user", content: "Hello" }]
});

const message = await client.messages.create({
  system: [
    { type: "text", text: systemPrompt, cache_control: { type: "ephemeral" } }
  ],
  tools: [
    { name: "tool1", ..., cache_control: { type: "ephemeral" } },
    { name: "tool2", ... },
    // ... 30 more tools
  ],
  messages: [...]
});

const message = await client.messages.create({
  system: [
    { type: "text", text: systemPrompt, cache_control: { type: "ephemeral" } }
  ],
  messages: [
    {
      role: "user",
      content: [
        {
          type: "text",
          text: ragContext,  // 30k token retrieved docs
          cache_control: { type: "ephemeral" }
        },
        { type: "text", text: userQuery }
      ]
    }
  ]
});

const message = await client.messages.create({
  system: [{ type: "text", text: systemPrompt, cache_control: { type: "ephemeral" } }],
  messages: [
    { role: "user", content: "Turn 1 query" },
    { role: "assistant", content: "Turn 1 response" },
    { role: "user", content: "Turn 2 query" },
    { role: "assistant", content: "Turn 2 response", cache_control: { type: "ephemeral" } },
    // ... cached up to turn 2
    { role: "user", content: "Turn 3 query" }  // New, not cached
  ]
});

cache_control: { type: "ephemeral" }

cache_control: { type: "ephemeral", ttl: "1h" }

// BAD: dynamic content trong system → invalidate cache
system: `You are bot. Current time: ${new Date()}. User: ${userId}.`
 
// GOOD: static system + dynamic in messages
system: `You are bot. ...`  // ← cacheable
messages: [
  { role: "user", content: `Time: ${new Date()}. UserID: ${userId}. Query: ${query}` }
]

// BAD: regenerate tools list mỗi request (object reference khác)
tools: regenerateTools(...)
 
// GOOD: const reference shared
const TOOLS = [...]; // Singleton
tools: TOOLS

// BAD: round-robin different user
processRequests([user1, user2, user1, user3, user2]); 
// → cache thrash
 
// GOOD: batch by user
processRequests([user1, user1, user1, user2, user2, user3]);
// → user1 hit cache 3 lần

// 8am cron: pre-warm before 9am peak
await client.messages.create({
  system: [{ type: "text", text: systemPrompt, cache_control: ... }],
  messages: [{ role: "user", content: "warmup" }],
  max_tokens: 10
});

// BAD: prefix 800 token < min → no cache
system: "Short system prompt 800 token..."
 
// GOOD: extend with boilerplate to > min
system: "Short system + brand voice + examples 3000 token..."

{
  "usage": {
    "input_tokens": 200,                  // Total input
    "cache_creation_input_tokens": 0,    // Write cost
    "cache_read_input_tokens": 5000,     // Read benefit
    "output_tokens": 500
  }
}

const hitRate = usage.cache_read_input_tokens / 
  (usage.cache_read_input_tokens + usage.cache_creation_input_tokens);

// BAD: random order each call
tools: shuffle(toolList)
 
// GOOD: sorted consistent order
tools: [...toolList].sort((a, b) => a.name.localeCompare(b.name))

// BAD: template substitution trong system
system: `Bot for ${tenantName}, day ${dayOfWeek}.`
 
// GOOD: parameterize via messages
system: "Bot per tenant + day"
messages: [{ role: "user", content: `Tenant: ${tenantName}, Day: ${dayOfWeek}. Query: ...` }]

// One-time script — không cần cache (no reuse)
node generate-report-once.js  // KHÔNG add cache_control

// Wiki content generation pipeline
const SYSTEM_PROMPT = `
You are VietCodex wiki writer...
[Brand voice 1000 token]
[Vietnamese style guide 1500 token]
[Cross-link strategy 1000 token]
[Schema requirements 800 token]
[FAQ format guide 500 token]
[Example output 700 token]
`;  // Total ~5500 token
 
const TOOL_DEFS = [
  { name: "search_wikimedia", ... },
  { name: "verify_url", ... },
  { name: "check_cross_link", ... },
  // ... 8 tools, ~3000 token total
];
 
// Generate 1 article
async function generateArticle(topic: string) {
  return client.messages.create({
    model: "claude-sonnet-4-5",
    system: [{ type: "text", text: SYSTEM_PROMPT, cache_control: { type: "ephemeral" } }],
    tools: [
      ...TOOL_DEFS.slice(0, -1),
      { ...TOOL_DEFS.at(-1), cache_control: { type: "ephemeral" } }  // breakpoint after tools
    ],
    messages: [
      { role: "user", content: `Write wiki article on: ${topic}` }
    ],
    max_tokens: 6000
  });
}
 
// Generate 20 article session
for (const topic of topics) {
  await generateArticle(topic);
}
 
// Result:
// Article 1: $0.015 (write) + $0.090 output = $0.105
// Article 2-20: $0.003 (read) + $0.090 output = $0.093 each
// Total 20 article: $0.105 + 19 × $0.093 = $1.873
 
// Without caching:
// Each article: 5500 × $3/M + 3000 × $3/M + 6000 × $15/M = $0.0165 + $0.009 + $0.090 = $0.1155
// Total: 20 × $0.1155 = $2.31
 
// Saving: $2.31 - $1.873 = $0.44 (19%)