DecisionRAG Uses Decision Models for Fast, Accurate Search
The trade is striking: instead of generating tokens, the model just ranks options per level, turning retrieval latency into a single forward pass per tree level. Teams building RAG pipelines over large corpora can use this to cut both query cost and latency without a separate reranker. It directly challenges dense-retrieval-plus-reranking stacks like the ones behind most production search systems today.
DecisionRAG replaces traditional retrieval and reranking with a decision model that scores all sibling options at once, navigating a hierarchical catalogue from broad categories down to specific passages. Tested against Wikipedia's 2.68 million passages, it completes precise lookups in 0.7 to 1.8 seconds at a cost of 0.2 to 2.3 US cents per query. Across multiple QA benchmarks, it outperforms conventional retrieval-plus-reranking pipelines by over ten percentage points in accuracy. It is also more than 30 times faster than letting a standard LLM reason step by step to find the answer.
DecisionRAG：用决策模型实现快速且精准的搜索
非常巧妙的思路，用Jev类决策模型来代替传统的搜索排序工具。这个方法借鉴了人类在图书馆查书的习惯，把庞大的资料库整理成大分类、文章标题到具体段落的层级结构。决策模型不需要逐字生成文本，只需在单次计算中对同级的所有选项同时打分，逐层筛选出最有价值的分支。
实际测试中，即使面对维基百科超过 268 万个段落的庞大资料库，系统也能在 0.7 到 1.8 秒内完成精准查找，单次查询成本只有 0.2 到 2.3 美分。在多个问答测试集上，这项技术不仅比传统搜索加重排方案的准确率高出十几个百分点，还比让普通大模型慢慢思考查找快了 30 多倍。
论文：https://t.co/kJEiEbpa4o
