Quality evaluation - Enterprise Data Platform

Change the pipeline, change the retrieval method, add documents - Every change can improve one question and break another. Quality evaluation helps you keep a standard question set for checking a KB, and track real retrieval quality from agents.

Two tabs

Tab	Used to
Golden question set	Manage questions along with the documents that must be retrieved
Live quality	Track real retrievals from agents

Golden question set tab

Element	Used to
Filter by KB	All KBs or a specific KB
Filter by result	All · Unchecked · Correct · Wrong
Import CSV	Import questions with their correct documents in bulk
Run full set	Run every question and score it
Add question	Add one question: the question text, an expected answer (optional) and the KB to check. A question counts as Correct when the KB’s top-K contains at least one of the selected documents

The quickest way to build golden questions: On the Retrieval test screen, click Add to golden question set after each test.

Live quality tab

Block	Meaning
Retrievals in 7 days	Number of agent calls, compared with the previous 7 days
No source found	Share of retrievals returning 0 results
Positive user ratings	Share of positive ratings agents send via the `submit_feedback` tool
Frequently unanswered questions	Knowledge gaps - A list of documents that should be added to the KB
Most problematic chunks	Chunks rated “not helpful” most often in 30 days
Quality by KB	Per KB: retrievals, average score, error rate over 30 days

Note: Check Frequently unanswered questions every week. It is the quickest way to find out which documents are still missing from a KB.