AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases
Researchers have introduced AgenticRAG, a novel framework designed to enhance retrieval and analysis within enterprise knowledge bases. Unlike standard Retrieval-Augmented Generation (RAG) pipelines that constrain language models to fixed candidate sets, AgenticRAG employs a lightweight harness over existing search infrastructure. This approach equips reasoning Large Language Models (LLMs) with autonomous tools for searching, finding, opening, and summarizing documents, enabling iterative information retrieval and evidence analysis. Benchmark tests demonstrate significant performance improvements, including a 49.6% recall@1 on BRIGHT, 0.96 factuality on WixQA, and 92% answer correctness on FinanceBench. Ablation studies highlight that shifting from single-shot retrieval to agentic tool use yields a 5.9x improvement. The system’s design, informed by pre-production deployments, proves suitable for real-world enterprise environments, offering superior quality and efficiency through multi-query search and in-document navigation capabilities.
Wire timeline
AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases
Researchers have introduced AgenticRAG, a novel framework designed to enhance retrieval and analysis within enterprise knowledge bases. Unlike standard Retrieval-Augmented Generation (RAG) pipelines that constrain language models to fixed candidate sets, AgenticRAG employs a lightweight harness over existing search infrastructure. This approach equips reasoning Large Language Models (LLMs) with autonomous tools for searching, finding, opening, and summarizing documents, enabling iterative information retrieval and evidence analysis. Benchmark tests demonstrate significant performance improvements, including a 49.6% recall@1 on BRIGHT, 0.96 factuality on WixQA, and 92% answer correctness on FinanceBench. Ablation studies highlight that shifting from single-shot retrieval to agentic tool use yields a 5.9x improvement. The system’s design, informed by pre-production deployments, proves suitable for real-world enterprise environments, offering superior quality and efficiency through multi-query search and in-document navigation capabilities.
cs.AI updates on arXiv.org