Experience-RAG Skill: Agent-Oriented Retrieval Orchestration
A new research paper introduces Experience-RAG Skill, an agent-oriented pluggable retrieval orchestration layer for retrieval-augmented generation systems. It analyzes tasks, consults experience memory, selects retrieval strategies, and returns structured evidence. On BeIR/nq, BeIR/hotpotqa, and BeIR/scifact, it achieves an nDCG@10 of 0.8924, outperforming fixed single-retriever baselines and competing with Adaptive-RAG-style routing. The work suggests encapsulating retrieval strategy selection as a reusable agent skill rather than hard-coding it.
Key facts
- Experience-RAG Skill is an agent-oriented pluggable retrieval orchestration layer.
- It positions between the agent and the retriever pool.
- It analyzes the current scene, consults experience memory, and selects retrieval strategies.
- Achieves nDCG@10 of 0.8924 on BeIR/nq, BeIR/hotpotqa, and BeIR/scifact.
- Outperforms fixed single-retriever baselines.
- Competitive with Adaptive-RAG-style routing.
- Published on arXiv with ID 2605.03989v1.
- Suggests retrieval strategy selection as a reusable agent skill.
Entities
Institutions
- arXiv