Model and prompt registry
A registry is the source of truth for what can be deployed. In LLM systems, that means more than model weights: prompts, retrieval indexes, tool schemas, guardrails, and eval results are part of the release.
What belongs in the registry
| Artifact | Why |
|---|---|
| Model or adapter | The learned component |
| Prompt template | The instruction and output contract |
| Retrieval config | Index version, chunking, embedding model, reranker |
| Tool schema | What the model can call |
| Eval report | Evidence that this version is safe to release |
| Approval / owner | Accountability and rollback contact |
The registry should point to immutable artifacts. Mutable "latest" names are fine for humans but dangerous as release dependencies.
Stages
Use explicit lifecycle stages: draft, candidate, staging, production, archived. Promotion should require passing the eval suite and recording the decision.
Rollback design
Every release should know its rollback target. For LLM apps, rollback may mean reverting a prompt while keeping the same model, or reverting a retrieval index while keeping the same prompt.
Pitfall
If prompt changes ship outside the registry, production behavior can drift without a model release. Treat prompts as code and as release artifacts.
Connects to: system prompts · vector indexes · CI/CD