Skip to content

What is semantic versioning for prompts and model configurations in production deployments?

Beginner · 2 min read · LLMOps & Production AI

Open in study mode to save this question and track progress

Short answer: Semantic versioning for prompts applies structured version tags (MAJOR.MINOR.PATCH) to prompt templates and model hyperparameter configurations, allowing engineering teams to audit, test, roll back, and correlate prompt changes directly with production telemetry.

Detail

Un-versioned prompt changes lead to silent quality regressions across production services.

Format: vMAJOR.MINOR.PATCH
- MAJOR (v2.0.0): Breaking changes (e.g. changing output schema from Markdown to JSON).
- MINOR (v1.1.0): New features / instructions (e.g. adding new guardrail rules or persona hints).
- PATCH (v1.0.1): Minor typos / wording tweaks (zero change to expected output structure).

Version Configuration Artifact Example

yaml
prompt_metadata:
  id: "customer_support_intent"
  version: "1.2.0"
  model: "gpt-4o-mini"
  temperature: 0.1
  max_tokens: 500
template: |
  You are an enterprise support agent. Classify user intent into {{ categories }}.

Example

Python prompt registry manager:

python
class PromptRegistry:
    def __init__(self):
        self.prompts = {}

    def register(self, name: str, version: str, template: str):
        key = f"{name}:{version}"
        self.prompts[key] = template

    def get(self, name: str, version: str) -> str:
        key = f"{name}:{version}"
        if key not in self.prompts:
            raise KeyError(f"Prompt '{key}' not found in registry.")
        return self.prompts[key]

registry = PromptRegistry()
registry.register("support", "1.0.0", "System rule v1...")
print("Loaded Prompt Version:", registry.get("support", "1.0.0"))

Interview tips

  • Discuss prompt registries (LangSmith, Phoenix, Humanloop) used to manage prompt deployment pipelines.
  • Explain associating prompt version tags with OpenTelemetry span traces.

Related Concepts

⬅ Back to LLMOps and Production AI · All topics