Advance Rag For Financial Information

Abstract

Financial documents are essential for understanding a companys performance, but their interpre- tation remains challenging due to their large volume, unstructured format, and domain-specific language. Traditional methods and even some Large Language Model (LLM)-based solutions often struggle with issues like hallucination, context loss, and difficulty processing complex ele- ments such as tables and charts. This project presents an advanced Retrieval-Augmented Genera- tion (RAG) based chatbot system specifically designed to overcome these limitations. The system takes financial documents in PDF format as input and leverages a pipeline that includes Query Enhancement, Semantic Chunking, and Reranking to optimize the accuracy and relevance of re- sponses. Using tools like Marker for PDF-to-Markdown conversion, LlamaIndex for parsing and context management, HuggingFace embeddings for semantic understanding, and Pinecone for ef- ficient vector storage, the system delivers accurate, context-aware answers to user queries. By integrating metadata and structuring document content into meaningful units, the proposed solu- tion significantly reduces hallucination, mitigates misinformation risks, and enhances the quality of information retrieval. The result is a robust and intelligent chatbot capable of assisting users in understanding complex financial documents with improved precision and reliability.