In today’s information-saturated financial world, surfacing in real time the nuggets that make a difference in the performance of companies, industries and economies is essential for the asset management community. Large Language Models (LLM) summarization is both unable to support real-time information and prone to hallucinations, leading to the emergence of Retrieval Augmented Generation (RAG), which focuses the attention of the LLM on specific fragments of recent information.
Most RAG systems leverage statistical distance between concepts through embeddings. We present here a different approach: Structured RAG (SRAG), which uses a symbolic approach to convert financial articles into readable structured data that guides the selection of text fragments for the prompt. The availability of this structure enables the detection of insights leveraging topic frequency, sentiment trends and intricate causal networks, producing Analytics Controlled Narratives that are accurate, relevant and rooted in analyzed data.