Per-Predictor Retry Configuration

[bdsaglam]

Problem

Currently, ActivityConfig applies to all LLM calls within a module uniformly. Different predictors may need different retry/timeout policies.

Example Use Case

class ResearchAgent(dspy.Module):
    def __init__(self):
        # Quick classification - short timeout, few retries
        self.classifier = dspy.Predict("text -> category")
        
        # Complex reasoning - long timeout, more retries
        self.researcher = dspy.ChainOfThought("topic -> detailed_analysis")

The classifier should fail fast, while the researcher needs patience.

Proposed Solution

Allow per-predictor configuration:

temporal_module = TemporalModule(
    agent,
    activity_config=ActivityConfig(timeout=60),  # default
    predictor_config={
        "classifier": ActivityConfig(timeout=10, max_retries=2),
        "researcher": ActivityConfig(timeout=300, max_retries=5),
    }
)

Implementation Notes

• Need to identify predictors by name or path
• Consider using module introspection (named_parameters())
• May require changes to how TemporalLM is created/configured

Priority

Medium