Exécution de notre outil de synthèse sur un nouvel article

Résumé : Retrieval-Augmented Generation (RAG) grounds language-model responses in externally retrieved evidence, yet weak retrieval or unsupported generation can still reduce reliability. This pilot study compares two lightweight safeguards: retrieval-confidence gating and post-generation evidence verification. Four configurations were evaluated on the same 30-question benchmark: Basic RAG, Confidence-Aware RAG, Verification-Based RAG, and a Combined approach. The benchmark used 30 short passages, 25 answerable questions, and five deliberately unanswerable questions. Retrieval employed normalized all-MiniLM-L6-v2 embeddings with FAISS top-3 search; generation and verification used Groq Compound Mini. All 120 runs completed without API errors. Basic and Verification-Based RAG achieved 30/30 preliminary correct outcomes. Confidence-Aware and Combined RAG achieved 29/30 because one answerable query received a similarity score of 0.261 and was rejected by the fixed 0.55 threshold. No unsupported final answer was observed. Mean response times were 0.50 s, 0.40 s, 1.29 s, and 0.95 s, respectively. The results reveal a practical trade-off: confidence gating can reduce model calls and encourage safe refusal, but an uncalibrated threshold may reduce answer coverage, while verification preserves coverage at additional latency.