Large language models (LLMs) excel at many tasks, from writing poems to summarizing articles. But how well do they handle complex word problems involving numerical reasoning within extensive texts? It turns out, not very well. Traditional LLMs struggle to simultaneously process vast amounts of text and perform accurate calculations. Simply prompting them to analyze and answer these problems results in low accuracy and high costs due to the sheer volume of tokens generated. However, a new research framework offers a clever workaround. This framework breaks down complex, number-heavy questions within long documents into four manageable steps: judging relevance, extracting key data, processing with code, and generating a final conclusion. Instead of overwhelming the LLM with the entire text, the framework uses smaller, more efficient models to filter out irrelevant sections. Then, a medium-sized model extracts relevant numerical data and organizes it into a structured table. The innovative step involves prompting the LLM to write code that analyzes this structured data. This sidesteps the LLM’s weakness in direct calculation and harnesses its strength in code generation. Finally, the LLM reviews the code's output and formulates a concise answer. This approach dramatically improves accuracy and significantly reduces processing costs by minimizing the use of large, expensive LLMs for basic filtering tasks. Tests on challenging benchmarks show this framework can effectively handle a wide variety of numerical reasoning tasks embedded within long texts. While not a universal solution, this framework represents a promising step towards making LLMs more effective and efficient problem-solvers, particularly in scenarios demanding both textual understanding and numerical reasoning.