Enterprise retail workflows are messy: they demand that an AI agent pull together evidence from product databases, order records, customer inquiries, and reviews—all in one reasoning chain. A new arXiv paper argues that this kind of heterogeneous chaining is exactly what agentic retrieval-augmented generation (RAG) should support, yet current benchmarks don't reflect it.
The paper notes that existing agentic RAG benchmarks rarely capture the complexity of mixing structured and unstructured sources in a single workflow. That gap matters because retail decisions often hinge on connecting a structured fact (e.g., an order status) with an unstructured signal (e.g., a customer complaint).
To move forward, the authors propose a research direction focused on building agentic RAG systems that can natively handle such mixed evidence. While the abstract stops short of detailing a specific benchmark or method, it frames the problem clearly: retail workflows are a natural testbed for agentic RAG, and current evaluation tools are not up to the task.