vault / agents / evals.md

Evaluation-driven development

Agents & LLMs 54 words 4 outgoing 4 backlinks

Compare relevance, retrieval recall, groundedness and latency against a baseline. Inspect examples as well as scores. In my current workflow, evaluation produces a report for engineers rather than an automatic deployment gate. That is a description of this system, not a rule that every team should avoid gates.

Read the full example: related post.

LINKS IN THIS NOTE

[[ Retrieval recall ]] Ask whether the expected sources reached the retrieved context. When the answer is wrong, this helps distinguish missing evidence from poor use of available evidence. The expected source set must itself be checked; a score against an incomplete reference can... [[ Groundedness ]] Check whether the answer's claims are supported by the retrieved sources. This differs from usefulness: an answer can stay within its evidence and still fail to answer the question. Read groundedness alongside relevance and retrieval recall, rather than reducing all... [[ Golden datasets ]] A golden set pairs questions with expected answers and source references. In the design described in the post, simulated personas propose document-anchored candidates, but people approve them before they count. Generated text is a proposal, not ground truth, and changing... [[ Online evaluation ]] Offline examples help compare changes under controlled conditions. Production feedback shows questions and failures the reference set missed. The two complement each other. Online signals need investigation: a thumbs-down tells us that something went wrong, not whether retrieval, content, configuration...