DeepSieve is designed with a dual-mode architecture of Naive and Graph: the Naive mode optimizes the processing speed of simple queries and is suitable for single data source scenarios, while the Graph mode adopts a graph structure to organize multi-source data relationships, which is more suitable for complex tasks that require cross-source reasoning (e.g., correlation analysis of academic literature). Users can switch modes through the environment variable RAG_TYPE. Both modes support the answer fusion function, but the Graph mode incurs higher computational overhead. The test data shows that Graph mode improves the accuracy on the Hotpot_QA dataset by about 181 TP3T over Naive mode.
This answer comes from the articleDeepSieve: a RAG Intelligent Information Screening Tool for Processing Complex Query SourcesThe
English 
简体中文 
日本語 
Deutsch 
Português do Brasil