{ "formula_1": { "strategy_comparison": { "k": 5, "full": { "accuracy": 0.5, "label": "Full schema" }, "bm25": { "accuracy": 0.485, "label": "BM25-only" }, "embedding": { "accuracy": 0.53, "label": "Embedding-only" }, "hybrid": { "accuracy": 0.545, "label": "Hybrid (RRF)" } }, "k_sweep_generation_accuracy": { "3": 0.47, "5": 0.545, "7": 0.47 }, "k_sweep_retrieval_recall": { "3": 0.53, "5": 0.71, "7": 0.83 }, "token_cost_reduction": { "full_schema_tokens": 2539, "avg_retrieved_tokens_k5": 1103.9, "reduction_pct": 0.565 }, "n_questions": 66, "n_tables": 13 }, "debit_card_specializing": { "accuracy": 0.433, "n_questions": 30, "n_tables": 5, "note": "No retrieval ablation run on this DB -- full schema only, used to demonstrate Stages A-C (core agent + self-correction + eval harness)." } }