Exécution de notre outil de synthèse sur un nouvel article

Résumé : Legal retrieval-augmented generation (RAG) can fail because the supporting authority was not retrieved, because retrieved authority was not selected as evidence, or because selected evidence did not produce an adequate answer. Treating these outcomes as one accuracy score obscures the intervention required. This study evaluates those failure sources on Legal RAG Bench, comprising 4,876 passages from the Victorian Criminal Charge Book and 100 expert-written questions with long-form reference answers and passage-level relevance labels. Four retrievers—BM25, word TF-IDF, character TF-IDF, and reciprocal-rank fusion (RRF)—were crossed with three deterministic evidence-selection policies in 12 conditions, producing 1,200 question-condition observations. BM25 achieved the highest exact annotated-passage Hit@5 (0.34), whereas RRF paired with Query-Focused selection produced the highest answer token F1 (0.2308). Under a strict hierarchy applied to all observations, 74.25% were annotated-passage retrieval misses, 9.92% were evidence-selection failures, 2.58% were answer-realization failures, and 13.25% were strictly supported successes. An evidence-constrained correction gate increased mean token F1 from 0.2043 to 0.2184 and ROUGE-L from 0.1425 to 0.1580. The question-paired token-F1 improvement was statistically significant (Wilcoxon p = .0213; mean difference = 0.0141, 95% bootstrap CI [0.0031, 0.0267]), although the lexical citation-support proxy decreased slightly from 0.9493 to 0.9448. The lexical groundedness proxy equaled 1.0000 because every answer consisted only of sentences selected from its cited passages. In the lowest lexical-overlap quartile, every retriever had Hit@5 = 0, identifying vocabulary mismatch as the principal unresolved bottleneck. The results support component-specific legal RAG evaluation, retrieval-aware correction, and abstention policies that expose residual risk.