11. Prompt Version Control

As prompts evolve through iteration, keeping track of changes becomes critical. Prompt versioning follows the same principles as code versioning but with unique challenges.

Prompt Storage Architecture

# prompts/customer-service/v2.3.yaml
version: "2.3"
created: "2024-03-15T14:32:00Z"
parent_version: "2.2"
model: "llama3:70b-instruct"
parameters:
  temperature: 0.7
  top_p: 0.9

template: |
  You are a customer service representative for Acme Corp.
  Policies:
  - Refunds processed within 5 business days
  - Escalation threshold: 3 prior contacts
  
  Customer query: {query}
  
  Response format:
  1. Acknowledge the issue
  2. Provide resolution timeline
  3. Offer alternatives if applicable

changes_from_2.2:
  - Added escalation threshold policy
  - Changed response format to numbered list

Git-Based Prompt Management

Store prompts in git with structured diffs:

# Commit prompt changes with semantic versioning
git add prompts/customer-service/v2.3.yaml
git commit -m "feat: add escalation threshold, restructure response"

# View prompt history
git log --oneline --follow prompts/customer-service/v2.3.yaml

# Diff specific versions
git diff v2.1..v2.3 -- prompts/customer-service/v2.3.yaml

A/B Testing with Versioned Prompts

import random
from prompty import PromptRegistry

registry = PromptRegistry("prompts/")

def serve_prompt(prompt_name, context, traffic_split=None):
    """Route traffic between prompt versions."""
    versions = registry.list_versions(prompt_name)
    
    if traffic_split:
        # Use specific version
        version = traffic_split.get("version", versions[-1])
    else:
        # Random split for A/B testing
        version = random.choice(versions)
    
    prompt = registry.load(prompt_name, version)
    return prompt.format(**context)

Rollback Strategy

When a new prompt version causes issues in production:

# Immediate rollback
kubectl set env deployment/prompt-service PROMPT_VERSION=v2.2

# Verify rollback
curl -X POST https://api.internal/health \
  -d '{"check": "prompt_response_quality"}'

Local verification checkpoint

Run the smallest example from this chapter in a local workspace and record the package version, runtime, data path, and observed output. If the result depends on model size, vector count, CPU/GPU backend, or available memory, note that constraint beside the exercise so the lesson remains reproducible.