检索质量

Relevance, coverage, freshness, and authorization all matter. A high similarity score alone does not prove that a passage answers the question.

主要要点

• Define relevance for the question being answered.
• Report cutoffs and labeling rules.
• Test freshness, permissions, and missing evidence.

Define relevance with the intended task in mind. A document about a product may be topically related but fail to answer a specific question about a version or date. Label examples of fully supporting evidence, partial evidence, and irrelevant material. Measure the candidate set and ranking separately. Recall at a chosen cutoff asks how much relevant material was retrieved; precision asks how much of the retrieved material is relevant. Rank-aware metrics assess whether the best evidence appears early. State the cutoff and labeling method with every score. Inspect failure patterns: exact identifiers missed by semantic search, synonyms missed by keyword search, outdated documents ranked above current ones, or passages cut away from their qualifications. Hybrid retrieval and reranking can help some cases, but must be evaluated on the same fixed examples. Include access restrictions and unanswerable queries in the test set. A system should not improve apparent relevance by returning unauthorized documents. When no adequate evidence exists, measure whether the application communicates that limitation instead of producing an unsupported answer.

1. In a constructed collection, four passages answer a question. A search returns five passages, of which three are relevant.
2. Precision at five is 3/5 = 60%; recall at five is 3/4 = 75%.
3. Inspect the missing relevant passage and the two irrelevant results before choosing a tuning change.

These invented counts show two different retrieval properties; neither alone measures final answer correctness.