Resolver Failing at: Extraction text must be a string, integer, or float.

[JoshuaSayagh]

Hey, once every 20 runs i would say, i get this error: Extraction text must be a string, integer, or float. i looked it up and it seems it comes from the LLM output being invalid? is there any known fix for this? or the only solution is to retry?

I saw a similar discussion about this unanswered.

Thanks for the help

[Yash-xoxo]

This usually happens when the LLM returns malformed JSON or unexpected types. A few things to try:

Quick fix:
Wrap your extraction in a retry loop with exponential backoff:

from tenacity import retry, stop_after_attempt, wait_exponential

@retry(stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=2, max=10))
def extract_with_retry():
    return your_extraction_call()

Root cause:
Check what the LLM is actually returning when it fails. Add logging:

try:
    result = extract(...)
except Exception as e:
    print(f"Raw LLM output: {raw_output}")
    raise

Often it's:

• null instead of a value
• Nested objects when expecting primitives
• Streaming artifacts

Better prompt:
Make your extraction schema stricter:

"Return ONLY a valid JSON object. All values must be strings, numbers, or booleans. Never return null."

What model are you using? Some are flakier with structured outputs than others.