ValueError on second iteration of loop - dspy.Predict

[mtp-pura]

This code loops through the first 's', returns a correct bool, but then fails on the next iteration:

lm = dspy.LM('ollama_chat/llama3.2:1b', api_base='http://localhost:11434', cache=False) 
dspy.configure(lm=lm)

# Define DSPY Signatures
classify= dspy.Predict('sentence -> sentiment: bool')
user = "worker"

# method for extracing 'sentences' 

# Analyse each sentence with the LM 
for s in sentences:
    question = f"Is \"{s}\" a task performed by a {user}?"
    current_sentence_response = classify(sentence=question)
    print(current_sentence_response.sentiment)

value = self.parse(signature, output, _parse_values=_parse_values)
return fn(instance, *args, **kwargs)
raise ValueError(f"Expected {signature.output_fields.keys()} but got {fields.keys()}")
ValueError: Expected dict_keys(['sentiment']) but got dict_keys([])```

With error handling I was able to check the history and the response it provides is:

Response:

I can't assist with this request.

[okhat]

Thanks for opening the issue! This seems like an LLM refusal. It doesn’t like your question.

You can wrap it in a try/except.

[mtp-pura]

Thanks. Yes, I now have some nested try/except going to manage it through staged questioning. The text I am feeding is quite fuzzy and unusual at times (like a project document's footer with obscure IDs and labels). Seems to be working okay at the moment, but still running into classic prompt "engineering" struggles. Will keep going as this is only day 1 with DSPy.

[okhat]

Awesome! Q: why are you concatenating a project document's footer into a string?

You can pass multiple fields into a module, by adjusting your signature.

instruction = "Is the supplied task something that this user can perform?"
module = dspy.Predict(dspy.Signature('task, user -> validity: bool', instruction))
module(task=..., user=...)

[mtp-pura]

Thanks for replying. That stacking of modules looks interesting. I may try to incorporate that.

I'll explain my use case but disclaimer: I am not a trained coder and I am more or less 'winging' it; so I am sure there are better ways to do everything I am trying to do.

My code is set up to :
Pull in PDFs from a folder. Convert the entire PDF to text. Separate by sentences (tokenize_string). Analyse each sentence (string) using a series of Predict and ChainOfThought modules. The footer is just caught up in the process during PDF handling.