Question

A financial services company deployed a RAG-based assistant that answers employee questions using an internal policy document store. During evaluation, reviewers notice that some answers sound plausible but contain claims that do not appear anywhere in the retrieved source documents. Which evaluation metric should the team prioritize to specifically detect this problem?

Accepted Answer

Answer groundedness (faithfulness), measuring whether the response is supported by the retrieved context

Answer

Retrieval latency, measuring how quickly the vector store returns candidate documents

Answer

Token throughput, measuring how many output tokens the model generates per second

Answer

Model parameter count, measuring the size of the underlying foundation model