Why AI Model Outputs Need a Versioned Schema Before Touching Your Database

A software developer writing for DEV Community argues that free or low-cost AI model outputs should be treated as untrusted upstream data, similar to a flaky webhook, rather than reliable function calls. The author describes a recurring failure where a summarizer app returned HTTP 500 errors because raw model completions containing markdown fences were passed directly to a JSON parser without validation. The proposed fix is a versioned Pydantic schema pinned between the model's response and the database write path, so malformed or drifted outputs are rejected with a 422 error before any row is touched. The article includes a Python code example using Pydantic's BaseModel to enforce strict field constraints on summary text, schema version, and confidence score. The author also recommends logging rejected completions as failed attempts, so on-call engineers retain a traceable record rather than facing a bare 500 error with no context.