Reproduce “Metadata-Driven Retrieval-Augmented Generation for Financial Question Answering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Michail Dadopoulos, Anestis Ladas, Stratos Moschidis, Ioannis Negkakis
A new multi-stage Retrieval-Augmented Generation (RAG) architecture enhances financial question answering by systematically integrating LLM-generated, multi-level metadata. The proposed system achieved a 31% relative improvement in F1-score compared to naive RAG, delivering robust performance with a custom, cost-effective reranker.
Key insights
- Direct integration of metadata into chunk embeddings ("contextual chunks") consistently improved final answer quality across different RAG architectures.
- A custom, metadata-aware reranker can achieve highly competitive performance, nearly matching commercial state-of-the-art solutions, at a fraction of the cost and with better transparency.
- The study demonstrated that blindly expanding retrieved chunks can harm performance, highlighting that focused, relevant context is more effective than increased but noisy information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.