Accuracy measures whether evaluated answers are correct under a specified criterion. In retrieval systems, correct answers depend on the retriever finding relevant content; a reader cannot recover an answer from documents it never receives. section 3.2.1
Keep these retrieval measures distinct:
- Accuracy assesses whether the model answers correctly; retrieval accuracy may instead count whether correct content appears within a specified result set. section 3.2.1 section 15.7.1
- Recall@k is binary: it counts a question as successful when a correct document appears anywhere in the top k. section 3.2.1
- MRR rewards higher placement by averaging the reciprocal rank of the first correct result. section 3.2.1
Compare accuracy with operating cost
On an 85-question evaluation set, baseline RAG reached 54.12% retrieval accuracy at 0.79 seconds per query. A-RAG reached 62.35% strict accuracy and 69.41% when semantically equivalent retrievals counted, but latency rose to 5.02 seconds. The gain matters only alongside the added cost. section 15.7.1 section 15.7.3
















































