CA-ThinkFlow: A RAG Framework for Chartered Accountancy

A novel framework known as CA-ThinkFlow has been developed to enhance the dependability of Large Language Models (LLMs) for intricate, jurisdiction-specific tasks, such as those in Indian Chartered Accountancy (CA). This parameter-efficient Retrieval-Augmented Generation (RAG) framework utilizes a 14B, 4-bit-quantized reasoning model, specifically the 14B-DeepSeek-R1, along with a layout-aware Docling extraction system that preserves the structure of documents during extraction. By employing a fundamental RAG approach, CA-ThinkFlow seamlessly integrates retrieved data into the prompt, leveraging the model’s inherent Chain-of-Thought (CoT) capabilities to establish context and generate responses. This innovation tackles the challenges LLMs face with multi-step numerical tasks and the complexities of legal regulations, while also addressing scalability in resource-limited environments.