Large language models (LLMs) are increasingly applied to question answering over user-provided documents, yet the influence of document encoding on their performance remains underexplored. This study evaluates how file format (TXT, DOCX, PDF, XML) affects latency and answer quality in a GPT-4–based system deployed in Microsoft Copilot Studio. Fifty scholarly articles (originally PDFs) were converted into each format, and two queries per article were executed across four GPT-4 agents configured with identical knowledge bases. Metrics included response time, answer length, source file size, and cross-format semantic consistency. Statistical tests revealed a significant effect of format on latency: XML produced the fastest responses, while file size showed no meaningful correlation. Longer answers modestly increased response time. Semantic content was largely preserved across formats, with average similarity scores above 0.91. These findings indicate that XML optimizes speed without compromising quality, whereas TXT, DOCX, and PDF deliver comparable semantic performance. The results provide practical guidance for deploying LLM-driven systems and highlight format selection as a critical design factor for efficient document-based question answering.