Standard RAG systems often become hallucination engines because naive PDF parsing destroys document structure. We solved this by implementing layout-aware partitioning using computer vision to identify headers and tables before extraction. By converting tables to structured HTML and utilizing a parent-child retrieval strategy, we preserved global context while maintaining surgical search precision. This engineering pivot improved our table query accuracy from 22% to 89%, proving that the real strength of enterprise AI lies in the fidelity of the ingestion pipeline rather than the model’s raw parameters.