Feedback Aggregation

For cases where argument specification names more than one value as an input, aggregation can be used.

Example

# Context relevance between question and each context chunk.
f_context_relevance = (
    Feedback(provider.context_relevance_with_cot_reasons, name = "Context Relevance")
    .on(Select.RecordCalls.retrieve.args.query)
    .on(Select.RecordCalls.retrieve.rets)
    .aggregate(numpy.mean)
)

The last line aggregate(numpy.mean) specifies how feedback outputs are to be aggregated. This only applies to cases where the argument specification names more than one value for an input. The second specification, for context was of this type.

The input to aggregate must be a method which can be imported globally. This function is called on the float results of feedback function evaluations to produce a single float.

The default is numpy.mean.