DeepSieve has a built-in three-level fault-tolerant processing mechanism: when the first retrieval of a sub-problem fails, the system analyzes the reason for failure (missing data/format mismatch) and automatically adjusts the retrieval strategy, with a maximum of two retries. The detailed process of each reflection is recorded in the query_{i}_results.jsonl log, which contains the original error, adjustment parameters, and secondary retrieval results. This mechanism improves the retrieval success rate on the Hotpot_QA dataset from the baseline 82% to 93%, and users can adjust the number of retries with the -max_reflexion_times parameter.
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