Structured Output Systems

A structured output system instructs the model to return data in a specific format — JSON, XML, or a custom schema. Downstream APIs parse and act on this output. When the format breaks, downstream systems break silently or loudly depending on how much validation exists.

System architecture

• LLM — generates output constrained to a schema
• Schema definition — JSON schema, Pydantic model, or TypeScript type
• Validation layer — parses and validates before passing downstream
• Downstream APIs — consume the structured output

What can go wrong

• Invalid JSON (model outputs trailing commas, unquoted keys)
• Schema mismatch (unexpected field names, wrong types)
• Partial outputs (model stops mid-object, context window exhausted)
• Extra text outside the schema (preamble, explanation, apology)
• Silent corruption (valid JSON, wrong values — harder to detect)

Detect

Reliai identifies:

• validation error rate spike
• schema mismatch patterns across model versions
• increase in output length without schema justification (sign of extra-schema text)
• downstream API error rate increase correlated with model changes

Understand

Incident example

An extraction pipeline begins failing downstream validation. Customer records are not being processed.

• Validation error rate: 0.3% → 14% over 30 minutes
• Trigger: model version update (gpt-4o-2024-05-13 → gpt-4o-2024-08-06)
• Impact: 2,800 records queued but unprocessable

Root cause

The new model version changed its formatting behavior for nested objects. It began including a brief explanation before the JSON block ("Here is the extracted data: {\n...}), which caused JSON parsers to fail on the prefix string.

The schema was unchanged. The prompt was unchanged. The model's output style shifted.

Reliai identified via:

• output comparison between old and new model version traces
• pattern detection on output prefix strings
• model version diff in deployment records

Fix

• Add output parser to strip non-JSON prefix text
• Add explicit prompt instruction: "Respond with only the JSON object. No explanation."
• Pin model version until behavior is validated

Prove

Key takeaway

Structured output failures are format integrity issues, not reasoning failures.

The model often knows the right answer. It fails to express it in the required format. Model version updates are the most common silent trigger — they change formatting behavior without changing task performance.