Neo4j-powered graph-rag system for financial insights on the Colombo stock exchange

Abstract

Financial annual reports contain rich but unstructured corporate information, making it difficult to efficiently extract relationships such as directors, subsidiaries, auditors, and ownership structures. Knowledge graphs provide a structured way to model these relationships and integrate heterogeneous information sources [1]. With recent advances in retrieval-augmented generation (RAG), graphbased retrieval can be combined with large language models to improve factual accuracy and context grounding in downstream analysis [4]. In this work, we transform Colombo Stock Exchange (CSE) annual reports into a Neo4j-based financial knowledge graph and integrate it with an LLM-driven Graph-RAG pipeline that supports naturallanguage financial querying. Our method introduces a focused retrieval strategy that extracts entities only from the most relevant text segments, overcoming LLM context limitations. This targeted approach enables near-complete entity capture and more accurate graph construction than full-document extraction.