Introduction: Standard Retrieval-Augmented Generation (RAG) often fails on multi-hop queries that require the synthesis of information from multiple sources. To address this limitation, we propose AtomRAG, a chain-of-prompts algorithm that decomposes a complex query into a series of atomic requirements for information retrieval. Materials and methods: This method utilizes a sequence of planning, parsing, and reasoning prompts to guide a large language model (LLM) in creating an explicit retrieval and synthesis plan before generating a final answer. We evaluated AtomRAG on FRAMES and BRIGHT benchmarks, comparing its performance against standard RAG and non-RAG baselines using an instruction-tuned model (Meta-Llama-3.1-8B-Instruct) and a reasoning model (DeepSeek-R1-Distill-Llama-8B). Results: Our results reveal a significant performance divergence. For the instruction-tuned model, AtomRAG improved accuracy by a relative 14.8% on the complex reasoning tasks in FRAMES. Conversely, the reasoning model proved incompatible with the instructive planning prompts, resulting in a performance degradation. On the less complex BRIGHT benchmark simpler methods performed better. Conclusions: We conclude that the effectiveness of AtomRAG is highly dependent on both query complexity and model architecture.
Igolnikov et al. (Thu,) studied this question.