| {"answerable": true, "category": "control", "metadata": {}, "question": "What does Recall at K measure in this benchmark?", "question_id": "ragbench-q001", "reference_answers": ["The fraction of labeled relevant chunks found in the first K results"], "relevant_chunk_ids": ["retrieval-basics-021fa30859::chunk-0000"]} | |
| {"answerable": true, "category": "semantic", "metadata": {}, "question": "Which retrieval metric gives more credit when the first useful evidence appears sooner?", "question_id": "ragbench-q002", "reference_answers": ["Reciprocal rank", "Mean Reciprocal Rank"], "relevant_chunk_ids": ["retrieval-basics-021fa30859::chunk-0000"]} | |
| {"answerable": true, "category": "lexical", "metadata": {}, "question": "Which model artifact version is approved for the canary pool?", "question_id": "ragbench-q003", "reference_answers": ["ms-2025-07-rc3"], "relevant_chunk_ids": ["model-serving-e5ce6a03a3::chunk-0002"]} | |
| {"answerable": true, "category": "lexical", "metadata": {}, "question": "What exact release identifier is the rollback target?", "question_id": "ragbench-q004", "reference_answers": ["ms-2025-06-stable"], "relevant_chunk_ids": ["model-serving-e5ce6a03a3::chunk-0002"]} | |
| {"answerable": true, "category": "acronym", "metadata": {}, "question": "What does the acronym RRF expand to?", "question_id": "ragbench-q005", "reference_answers": ["Reciprocal Rank Fusion"], "relevant_chunk_ids": ["hybrid-rrf-4479d7429d::chunk-0000"]} | |
| {"answerable": true, "category": "paraphrase", "metadata": {}, "question": "Why does the hybrid method merge rank positions instead of averaging lexical and cosine values?", "question_id": "ragbench-q006", "reference_answers": ["BM25 and cosine scores have unrelated numeric scales"], "relevant_chunk_ids": ["hybrid-rrf-4479d7429d::chunk-0001"]} | |
| {"answerable": true, "category": "control", "metadata": {}, "question": "How many candidates does each component retrieve before reciprocal rank fusion?", "question_id": "ragbench-q007", "reference_answers": ["Twenty", "20"], "relevant_chunk_ids": ["hybrid-rrf-4479d7429d::chunk-0002"]} | |
| {"answerable": true, "category": "semantic", "metadata": {}, "question": "After L2 normalization, what familiar similarity measure is equivalent to the vector dot product?", "question_id": "ragbench-q008", "reference_answers": ["Cosine similarity"], "relevant_chunk_ids": ["dense-embeddings-7a474808ec::chunk-0001"]} | |
| {"answerable": true, "category": "control", "metadata": {}, "question": "What embedding width does all-MiniLM-L6-v2 produce?", "question_id": "ragbench-q009", "reference_answers": ["384 dimensions", "384-dimensional embeddings", "384"], "relevant_chunk_ids": ["dense-embeddings-7a474808ec::chunk-0000"]} | |
| {"answerable": true, "category": "lexical", "metadata": {}, "question": "Which opaque code beginning with ZXQ is given as difficult for semantic vectors?", "question_id": "ragbench-q010", "reference_answers": ["ZXQ-417"], "relevant_chunk_ids": ["dense-embeddings-7a474808ec::chunk-0002"]} | |
| {"answerable": true, "category": "distractor", "metadata": {}, "question": "Which probe determines whether a serving instance should receive traffic?", "question_id": "ragbench-q011", "reference_answers": ["Readiness probe"], "relevant_chunk_ids": ["model-serving-e5ce6a03a3::chunk-0001"]} | |
| {"answerable": true, "category": "paraphrase", "metadata": {}, "question": "What request-grouping technique trades a little queueing delay for better accelerator throughput?", "question_id": "ragbench-q012", "reference_answers": ["Dynamic batching"], "relevant_chunk_ids": ["model-serving-e5ce6a03a3::chunk-0000"]} | |
| {"answerable": true, "category": "acronym", "metadata": {}, "question": "What TRACE identifier connects retrieval and generation spans?", "question_id": "ragbench-q013", "reference_answers": ["TRACE-RAG-8841"], "relevant_chunk_ids": ["observability-6ae870f273::chunk-0000", "observability-6ae870f273::chunk-0001"]} | |
| {"answerable": true, "category": "semantic", "metadata": {}, "question": "Why should a RAG service time retrieval, context assembly, and generation separately?", "question_id": "ragbench-q014", "reference_answers": ["To identify which stage caused a slowdown"], "relevant_chunk_ids": ["observability-6ae870f273::chunk-0000"]} | |
| {"answerable": true, "category": "control", "metadata": {}, "question": "What computation is used for exact vector search on the small evaluation corpus?", "question_id": "ragbench-q015", "reference_answers": ["NumPy matrix multiplication", "A NumPy matrix multiplication"], "relevant_chunk_ids": ["vector-indexing-21fd7fc623::chunk-0000"]} | |
| {"answerable": true, "category": "acronym", "metadata": {}, "question": "What does HNSW stand for?", "question_id": "ragbench-q016", "reference_answers": ["Hierarchical Navigable Small World", "Hierarchical Navigable Small World graphs"], "relevant_chunk_ids": ["vector-indexing-21fd7fc623::chunk-0001", "vector-indexing-21fd7fc623::chunk-0002"]} | |
| {"answerable": true, "category": "distractor", "metadata": {}, "question": "Why is redacting output after generation insufficient for document access control?", "question_id": "ragbench-q017", "reference_answers": ["Unauthorized chunks must be prevented from entering the context"], "relevant_chunk_ids": ["security-privacy-aa437fce64::chunk-0000"]} | |
| {"answerable": true, "category": "lexical", "metadata": {}, "question": "Which synthetic access-control policy marker appears in the corpus?", "question_id": "ragbench-q018", "reference_answers": ["SEC-POL-009"], "relevant_chunk_ids": ["security-privacy-aa437fce64::chunk-0001", "security-privacy-aa437fce64::chunk-0002"]} | |
| {"answerable": true, "category": "semantic", "metadata": {}, "question": "How can using too much overlap harm the retrieval ranking?", "question_id": "ragbench-q019", "reference_answers": ["Near-duplicate candidates can crowd the ranking"], "relevant_chunk_ids": ["chunking-62e846459e::chunk-0001"]} | |
| {"answerable": true, "category": "paraphrase", "metadata": {}, "question": "When should relevance annotations be reviewed again?", "question_id": "ragbench-q020", "reference_answers": ["When source text or chunk configuration changes"], "relevant_chunk_ids": ["chunking-62e846459e::chunk-0002"]} | |
| {"answerable": true, "category": "paraphrase", "metadata": {}, "question": "Which four inputs make an experiment metric auditable?", "question_id": "ragbench-q021", "reference_answers": ["Code, configuration, data versions, and model identity"], "relevant_chunk_ids": ["mlops-reproducibility-7e14c23dfe::chunk-0000"]} | |
| {"answerable": true, "category": "distractor", "metadata": {}, "question": "Should a build timestamp be allowed to change content IDs or ranking tie order?", "question_id": "ragbench-q022", "reference_answers": ["No"], "relevant_chunk_ids": ["mlops-reproducibility-7e14c23dfe::chunk-0001"]} | |
| {"answerable": false, "category": "unanswerable", "metadata": {}, "question": "Which GPU accelerator model was used to produce the benchmark results?", "question_id": "ragbench-q023", "reference_answers": [], "relevant_chunk_ids": []} | |
| {"answerable": false, "category": "unanswerable", "metadata": {}, "question": "What neural reranker is enabled by default after hybrid fusion?", "question_id": "ragbench-q024", "reference_answers": [], "relevant_chunk_ids": []} | |