Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources
- Published
- Source
- arXiv
- Paper number
- 1356
- Field
- RAG / Agents
- arXiv ID
- 2610.25096
Key points
- A recursive composition pipeline built 1,680 multi-source queries across three difficulty levels on top of a retail database created from simulated seller-buyer interactions.
- A domain expert with over ten years of experience of over 10 years marked 90% of the queries as questions a seller-side analyst would ask in practice, including 97.5% at Level-1, 90.0% at Level-2, and 82.5% at Level-3.
- RetailRAG reached an average LLM-judge score of 0.429, above 0.293 for the strongest baseline HetaRAG, while using 7.45 API calls per query against HetaRAG's 25.21.
- Among ten LLM backbones, even the strongest Qwen3.6-Plus and GPT-5.4 reached only 0.511 and 0.508 average LLM-judge scores, leaving RetailBridge unsolved.
- Every backbone lost about half or more of its Level-1 performance by Level-3, and Text-only queries collapsed the steepest, making dense text retrieval the dominant modality bottleneck.
Paper links
External research summaries. These are not HDATF publications or measured product results.