We present FinLLaMA-RAG, a Retrieval-Augmented Generation (RAG) framework built on LLaMA 3 to address complex document-level question answering. The model integrates dense retrieval, contextual fusion, and multi-hop reasoning to extract accurate, coherent answers from long and cross-referenced documents. Designed with high-stakes applications in mind, FinLLaMA-RAG supports automated tax compliance, financial fraud investigation, and regulatory reporting. Through joint optimization of retrieval and generation, it achieves strong results across several financial QA datasets. This work contributes to the development of explainable and reliable AI tools for financial and policy-oriented document analysis.

Text-attributed graph node classification is still a challenge since it needs to reason about the topology structure simultaneously with the free-text semantics. Although graph neural network can perform well on structural propagation,they tend to be blind for the details in the text associated with nodes. On the other hand, LLMs have excellent NLU skills and are weak on structured,multi-hop reasoning over network agents.To address the above gap, in this work we propose FinSCRA, a novel LLM-powered multi-chain reasoning framework to inject domain-aware reasoning capability into a financial LLM with parameters efficient fine-tuning. Specifically, our framework designs a hierarchy of structured reasoning chains (single-hint,parallel, cascaded, and hybrid methods to extract and fuse the semantic signals like sentiment, correlation, and risk signals in the nodes’ text.A fusion layer based on fuzzy logic fuses the results of different reasoning lines for better robustness and explainability.While FinSCRA is generic and can be applied to other types of text-attributed graphs, here we assess its performance on credit risk analysis in supply chain networks,on the task of entity relation extraction, in which entities are related through their financial relation and described with rich text reports; we show experimentally on realworld datasets that our model FinSCRA greatly outperforms graphbased as well as LLM-based baselines,as an accurate and explainable technique to perform node classification over complex networked systems.We release our code and models for further research on LLM-grap.

This paper presents a novel Retrieval-AugmentedGeneration (RAG) framework tailored for complex questionanswering tasks, addressing challenges in multi-hop reasoningand contextual understanding across lengthy documents. Builtupon LLaMA 3, the framework integrates a dense retrievalmodule with advanced context fusion and multi-hop reasoningmechanisms, enabling more accurate and coherent responsegeneration. A joint optimization strategy combining retrievallikelihood and generation cross-entropy improves the model’srobustness and adaptability. Experimental results show that theproposed system outperforms existing retrieval-augmented andgenerative baselines, confirming its effectiveness in deliveringprecise, contextually grounded answers.

This paper presents a novel Retrieval-Augmented Generation (RAG) framework tailored for complex question answering tasks, addressing challenges in multi-hop reasoning and contextual understanding across lengthy documents. Built upon LLaMA 3, the framework integrates a dense retrieval module with advanced context fusion and multi-hop reasoning mechanisms, enabling more accurate and coherent response generation. A joint optimization strategy combining retrieval likelihood and generation cross-entropy improves the model's robustness and adaptability. Experimental results show that the proposed system outperforms existing retrieval-augmented and generative baselines, confirming its effectiveness in delivering precise, contextually grounded answers.

This paper presents a novel Retrieval-AugmentedGeneration (RAG) framework tailored for complex questionanswering tasks, addressing challenges in multi-hop reasoningand contextual understanding across lengthy documents. Builtupon LLaMA 3, the framework integrates a dense retrievalmodule with advanced context fusion and multi-hop reasoningmechanisms, enabling more accurate and coherent responsegeneration. A joint optimization strategy combining retrievallikelihood and generation cross-entropy improves the model’srobustness and adaptability. Experimental results show that theproposed system outperforms existing retrieval-augmented andgenerative baselines, confirming its effectiveness in deliveringprecise, contextually grounded answers.