DSPy

pip install "llm-beacon[dspy]"

import dspy
from typing import Literal
from beacon.integrations.dspy import BeaconJSONAdapter, BeaconValidateSchema

dspy.configure(lm=dspy.LM("openai/gpt-4o"), adapter=BeaconJSONAdapter())

class IntentSignature(dspy.Signature):
    """Extract user intent from a message."""
    message: str = dspy.InputField()
    intent: Literal["cancel", "refund", "question"] = dspy.OutputField()
    confidence: float = dspy.OutputField()

class IntentModule(dspy.Module):
    def __init__(self):
        super().__init__()
        self.extract = dspy.ChainOfThought(IntentSignature)
        self.validate = BeaconValidateSchema(IntentSignature)

    def forward(self, message):
        return self.validate(self.extract(message=message))

`BeaconJSONAdapter`

A subclass of dspy.JSONAdapter that keeps DSPy’s prompt formatting but parses responses with extract_json(kind="object"):

strict JSON only, no repair;
missing or malformed fields raise DSPy’s AdapterParseError, which feeds DSPy’s own retry and fallback logic;
values are then converted to the field annotations by DSPy’s parse_value (that conversion is DSPy’s behaviour, not beacon’s).

`BeaconValidateSchema`

Builds a JSON Schema from a signature’s output fields (via Pydantic) and validates predictions against it, raising SchemaValidationError. A JSON Schema mapping works too:

BeaconValidateSchema({"type": "object", "properties": {"n": {"type": "integer"}}, "required": ["n"]})

Why it matters for optimization

DSPy optimizers (e.g. BootstrapFewShot) compare metrics across many runs. With a deterministic parser and validator, a metric change comes from your prompts and demos, never from flaky parsing.