01
Calculate precision and recall for the sample
The top-three list contains two relevant documents, D2 and D4, and one nonrelevant document, D1. Precision is the relevant retrieved count divided by the retrieved count: two divided by three. Recall is the relevant retrieved count divided by all relevant documents in the key: also two divided by three. Both are approximately 66.7 percent here, but they have different denominators. D6 is the missed relevant document. Keep the identifiers beside the fractions so the equal percentages do not hide the difference between the two questions.
02
Compare a longer candidate list
Now extend the fictional ranking to five candidates: D2, D1, D4, D3 and D6. All three relevant documents are present, so recall becomes three divided by three, or 100 percent. Precision becomes three divided by five, or 60 percent. The longer list improves recall in this example while reducing precision. This is a consequence of these particular labels and candidates, not a universal performance result. The sixth document remains outside the list, and no conclusion about answer quality follows from the candidate counts alone.
03
Record relevance and support separately
Use fields for question identifier, collection version, relevance key, ordered candidates, cutoff, relevant hits, precision and recall. Add a separate review field for whether any proposed answer is supported by quoted source material. The retrieval worksheet should not fill that answer-support field automatically from its percentages. A complete record can therefore show strong recall while leaving answer support unassessed. When revising the relevance key, preserve its earlier version so a later reviewer can distinguish a changed judgement from a changed retrieval result.