LLMQuoter: Enhancing RAG Capabilities Through Efficient Quote Extraction From Large Contexts
This provides a scalable and practical solution for researchers and practitioners to enhance RAG capabilities without extensive retraining, though it appears incremental as it builds on existing distillation and fine-tuning methods.
The paper tackles the problem of extracting relevant textual evidence for Retrieval Augmented Generation (RAG) by introducing LLMQuoter, a lightweight model that uses a 'quote-first-then-answer' strategy, achieving over 20-point accuracy gains compared to full-context approaches like RAFT.
We introduce LLMQuoter, a lightweight, distillation-based model designed to enhance Retrieval Augmented Generation (RAG) by extracting the most relevant textual evidence for downstream reasoning tasks. Built on the LLaMA-3B architecture and fine-tuned with Low-Rank Adaptation (LoRA) on a 15,000-sample subset of HotpotQA, LLMQuoter adopts a "quote-first-then-answer" strategy, efficiently identifying key quotes before passing curated snippets to reasoning models. This workflow reduces cognitive overhead and outperforms full-context approaches like Retrieval-Augmented Fine-Tuning (RAFT), achieving over 20-point accuracy gains across both small and large language models. By leveraging knowledge distillation from a high-performing teacher model, LLMQuoter achieves competitive results in a resource-efficient fine-tuning setup. It democratizes advanced RAG capabilities, delivering significant performance improvements without requiring extensive model retraining. Our results highlight the potential of distilled quote-based reasoning to streamline complex workflows, offering a scalable and practical solution for researchers and practitioners alike.