Semantic Retrieval and Fine-Tuned T5-Small for Investor Query Answering from Annual Reports

Abstract

The complexity and volume of corporate annual reports often make it challenging for investors to extract key financial insights. This research presents a semantic question-answering system designed to simplify this process. The contribution of this paper is an end-to-end system that combines a fine-tuned T5-small model with Facebook AI Similarity Search (FAISS)-based semantic retrieval to deliver context-aware answers to investor queries from annual reports. It processes corporate annual reports, retrieves relevant sections, and generates real-time structured responses. The system is deployed through a user-friendly Streamlit interface, ensuring seamless user interaction. This paper introduces an end-to-end financial question-answering pipeline that combines semantic retrieval and transformer-based answer generation, which reduces the manual effort in investment analysis.