AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases
AgenticRAG, a novel agentic harness designed for retrieval and analysis within enterprise knowledge bases, is detailed in arXiv:2605.05538. This system mitigates the reliance of traditional RAG pipelines on search stacks by introducing a lightweight framework that integrates with existing enterprise search systems. It empowers a reasoning LLM with capabilities for searching, finding, opening, and summarizing, thus facilitating iterative retrieval, document exploration, and independent evidence evaluation. In three open benchmarks, AgenticRAG demonstrated impressive results: 49.6% recall@1 on BRIGHT (an increase of 21.8 pp over the leading embedding baseline), 0.96 factuality on WixQA (a 13% relative enhancement), and 92% answer accuracy on FinanceBench (within 2 pp of oracle access). Ablation studies reveal that the primary factor is the change in retrieval approach.
Key facts
- AgenticRAG is a practical agentic harness for retrieval and analysis over enterprise knowledge bases.
- It reduces overdependence on the search stack by layering a lightweight harness on top of existing enterprise search infrastructure.
- The system equips a reasoning LLM with search, find, open, and summarize tools.
- It enables iterative retrieval, document navigation, and autonomous evidence analysis.
- On BRIGHT benchmark: 49.6% recall@1 (+21.8 pp over best embedding baseline).
- On WixQA: 0.96 factuality (+13% relative improvement).
- On FinanceBench: 92% answer correctness (within 2 pp of oracle access).
- Ablation studies show the most significant factor is the shift in retrieval strategy.
Entities
—