Send 100 requests to an LLM using OpenRouter

  • Send 100 requests to the same LLM using OpenRouter
    • Sample code
    • OpenRouter allows you to reuse the code and only change the model name
  • How many have pure JSON
    • Sometimes the response is Markdown JSON which is still valid JSON once you strip out the Markdown annotation
    • Some LLMs like Claude preface the answer with “Here is the information you requested” 🙂
  • How many contain valid JSON
    • you can use regular expressions to extract JSON out of the string response
    • and often it will be valid JSON even if the response itself isn’t pure JSON
  • How many have valid schema?
    • then check to see if the JSON complies with the schema provided in the input
  • How many extract the accurate value?
    • only after doing all these steps, you will be able to calculate if the extracted value is accurate
  • First, calculate the Structured Output Percentage Stats
    • not all LLMs behave the same way