{
  "coverage_notes": [
    "The TechQA track is external public technical-support data and is reported separately from the controlled synthetic operations benchmark.",
    "The tracked sample is deterministic so CI, Streamlit, and GitHub Pages can reproduce the public-data metrics without downloading the upstream dataset.",
    "Provider-backed embedding comparison remains optional and is not claimed until a credentialed run publishes a separate result."
  ],
  "report_type": "techqa_public_benchmark_profile",
  "summary": {
    "answerable_case_count": 384,
    "answerable_context_coverage": 1.0,
    "average_contexts_per_case": 0.8,
    "benchmark_track": "external_public_rag",
    "dataset": "nvidia/TechQA-RAG-Eval",
    "failed_case_count": 197,
    "failure_rate": 0.4104,
    "impossible_abstention_rate": 0.1146,
    "impossible_case_count": 96,
    "impossible_case_share": 0.2,
    "license": "Apache-2.0",
    "provider_backed_embedding_result_published": true,
    "retrieval_hit_rate_at_3": 0.8073,
    "sample_case_count": 480,
    "sample_path": "data/public/techqa_rag_eval_sample.jsonl",
    "sample_scope": "tracked_compact_public_sample",
    "sample_selection_policy": "first_answerable_plus_first_impossible_20pct",
    "source_url": "https://huggingface.co/datasets/nvidia/TechQA-RAG-Eval",
    "status": "evaluated",
    "top1_citation_accuracy": 0.724,
    "tracked_sample_default_limit": 480,
    "unique_document_count": 337
  }
}