Prompt Versioning ve A/B Testing

# prompts/customer-support.yaml
metadata:
  name: customer_support_bot
  version: "1.4.2"
  author: "yusuf@sukruyusufkaya.com"
  changelog:
    "1.0.0": Initial release (2026-01-15)
    "1.1.0": Added "be more concise" instruction (2026-02-03)
    "1.2.0": Added prompt injection defense (2026-02-15)
    "1.3.0": Switched to JSON output (2026-03-01)
    "1.4.0": Added Turkish-specific tone guidance (2026-04-10)
    "1.4.1": Fixed: model was being overly formal (2026-04-12)
    "1.4.2": Tweak: pasif/aktif voice balance (2026-04-22)
 
system_prompt: |
  Sen bir teknik destek asistanısın.
  ...
 
variables:
  - name: customer_tier
    type: enum
    values: [free, plus, pro, enterprise]
    description: Müşterinin abonelik seviyesi
 
  - name: history
    type: array
    description: Önceki sohbet geçmişi (son 5 mesaj)
 
response_format:
  type: json_schema
  schema_path: ./schemas/support_response.json
 
evaluation:
  test_cases: ./tests/customer_support.jsonl
  metrics:
    - accuracy
    - tone_match
    - safety_violations
  thresholds:
    accuracy: ">= 0.92"
    tone_match: ">= 4.0/5"
    safety_violations: "== 0"

// Basit A/B router
async function getPromptVersion(userId) {
  // Hash-based deterministic routing — aynı kullanıcı hep aynı versiyona düşer
  const hash = simpleHash(userId);
  const bucket = hash % 100;
 
  if (bucket < 10) {
    return "v2.0.0-beta"; // %10 trafik
  } else {
    return "v1.4.2"; // %90 trafik
  }
}
 
async function answerCustomer(userId, question) {
  const version = await getPromptVersion(userId);
  const prompt = await loadPrompt(version);
  const response = await callLLM(prompt, question);
 
  // Telemetri kaydet
  await logTelemetry({
    userId,
    promptVersion: version,
    questionLength: question.length,
    responseLength: response.length,
    timestamp: Date.now(),
  });
 
  return response;
}