Empirical Evaluation of PDF Parsing and Chunking for Financial Question Answering with RAG

This project is a classic empirical study accompanying a research paper. While it provides valuable insights into how different parsing strategies affect financial QA performance, it lacks a technical moat. The repository currently shows 0 stars and 8 forks, which, given its 4-day age, suggests it is likely being used by a small research circle or the authors themselves. Defensibility is low because the findings are easily reproducible, and the 'code' is a set of scripts for benchmarking existing tools (like Unstructured or LlamaParse) rather than a novel engine. From a frontier-lab perspective, this space is high-risk; OpenAI, Google, and Anthropic are rapidly improving native multimodal capabilities (e.g., GPT-4o, Gemini 1.5 Pro) that ingest PDFs directly via vision or native OCR, bypassing the need for complex external chunking and parsing pipelines. Furthermore, 'Long Context' windows (1M+ tokens) are beginning to make the RAG chunking strategies evaluated here less critical for single-document analysis. Competitively, it sits in a crowded space with established players like Unstructured.io and LlamaIndex who are institutionalizing these benchmarks into production-grade SDKs.