Reflection

After a run, chain.reflect(...) asks an LLM to analyse how well the chain accomplished the task — surfacing weak steps and concrete suggestions.

result = chain.execute(context)

reflection = chain.reflect(
    task_description="Analyze customer sentiment and extract key themes",
)
print(reflection)

Parameter	Type	Default	Purpose
task_description	str	— (required)	The original goal to judge against.
context	ReasoningContext | None	None	Defaults to the last execution context.
language	Language | None	None	Defaults to the context language.
options	ReflectionOptions | None	None	Controls what goes into the prompt.

It raises RuntimeError if no run has happened yet.

Options

ReflectionOptions tunes the reflection prompt:

from mmar_carl import ReflectionOptions

chain.reflect(
    task_description="...",
    options=ReflectionOptions(include_dependency_analysis=False),
)

When include_metric_scores=True (default), any metric scores are fed into the prompt as concrete quality signals — e.g. a low keyword_coverage on a step nudges the model to suggest better step_context_queries.

Reflection over a dataset

Pair reflection with DatasetEvaluator: a SelectionStrategy (ThresholdStrategy / TopKWorstStrategy) picks the worst-scoring cases so you reflect on the failures that matter.

See also

• Metrics · Datasets & evaluation
• Reflection example in the repo.