Scholarlink

This article proposes RAG-EVO (Evolutionary, Self-Improving RAG Agent), an adaptive generation-enhanced retrieval architecture that incorporates heuristic introspection mechanisms, persistent vector memory and evolutionary learning through iterative logs. The technique was evaluated in a simulated scenario using real legal-epidemiological data and compared with approaches such as Self-RAG, HyDE, Multi-Query RAG, MMR, and ReAct. RAG-EVO showed superior performance in factual consistency and completeness, achieving a composite accuracy score of 92.6%. The architecture is particularly suitable for domains that require robustness in the accuracy of the answers generated through LLMs, traceability and continuous adaptation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RAG-EVO: Increasing the Reliability and Autonomy of LLMs via Iterative Recovery

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Diego Rodrigues,
  • Daniela Mascarenhas de Queiroz Trevisan,
  • Sílvia Araújo,
  • David Nadler Prata

摘要

This article proposes RAG-EVO (Evolutionary, Self-Improving RAG Agent), an adaptive generation-enhanced retrieval architecture that incorporates heuristic introspection mechanisms, persistent vector memory and evolutionary learning through iterative logs. The technique was evaluated in a simulated scenario using real legal-epidemiological data and compared with approaches such as Self-RAG, HyDE, Multi-Query RAG, MMR, and ReAct. RAG-EVO showed superior performance in factual consistency and completeness, achieving a composite accuracy score of 92.6%. The architecture is particularly suitable for domains that require robustness in the accuracy of the answers generated through LLMs, traceability and continuous adaptation.