Xiaohongshu's new data-retrieval system hits 96.6% accuracy

Enterprise analytics agents often fail to find the right data assets, returning wrong Hive tables and missing usage context that leads to metric misinterpretation. Xiaohongshu researchers Mengdi Chen, Yuanxin Huang, Yulin Jiang, and Wei Sun identified four root causes: semantic gaps, entity ambiguity, schema drift, and asset-usage gaps. They built a two-layer solution deployed in Xiaohongshu's commercial advertising data warehouse, which holds 5,300+ Hive tables across 14 domains.

The first layer is a three-tier dual-purpose knowledge base (179 documents with an eight-section annotation template) that serves both retrieval and generation, kept fresh via a closed-loop refresh pipeline with day-level freshness and 30-second hot-reload. The second layer uses a Graph-Guided Retriever (GGR) with a 2,859-node knowledge graph as a candidate gate and intent routing, delivering a 71.6x token reduction. A Scene-Aware Ranker (SAR) applies 19-class entity recognition and explicit scenario annotations, with negative knowledge alone contributing 25 percentage points of Hit@10 gain. On two 100-question benchmarks, Hit@10 rose from 19.1% to 96.6% and knowledge coverage from 56% to 77%, at 4.84–5.33s end-to-end latency.