Skip to main content

Logging & Recovery

Resilient prompts need observability. prompt-fn layers structured logging, fallbacks, and manual content parsing so you can still ship when a provider emits malformed JSON.

pino logger​

import pino from 'pino';

const logger = pino({ name: 'Adder Prompt', level: 'debug' });

const addNumbers = definePrompt({
  ...,
  logger,
});
  • Each invocation logs module, errorName, and content when something goes wrong.
  • Use logger.child({ module: name }) if you need per-prompt context.

Recovery flow​

  1. We call generateObject / generateText with outputSchema.
  2. If generation.output is missing, we inspect generation.content and attempt to parse JSON ourselves.
  3. recoverFromContent (see src/utils/outputRecovery.ts) strips fences, trims whitespace, and re-validates.
  4. If manual parsing works, we log a warning instead of an error and return the recovered payload.
  5. If everything fails, the original AI SDK error bubbles up, still enriched with logger metadata.

You can override the recovery strategy by wrapping recoverFromContent or by transforming generation.content before calling outputSchema.parse.

Tuning provider requests​

  • Ollama OpenAI-compatible endpoints often wrap JSON in quotes. The built-in recovery handles this scenario automatically.
  • For providers that support structured output mode (e.g., Gemini 1.5, OpenAI Responses API), keep outputSchema strict—prompt-fn already passes the JSON schema to the provider.
  • If you rely on generateText, consider supplying a lightweight outputSchema and falling back to recoverFromContent for consistency.