Agents read documents through a PDF library and a search index. A failure on that path costs twice: compute spent on pages that did not need an expensive rasterization and OCR step, and wrong answers the agent passes on without noticing. The talk discusses how you can audit your ingestion pipeline with small experiments that can be adopted to your production environments. Why agents? A human retries a bad search result. An agent quotes it and hands it to the next step. A lack of rigor in the first steps of your pipeline make it very hard to attribute errors let alone fixing them in the generation step. Follow the money: → Reading a PDF is based on heuristics. We will learn how many percent of pages that are parsed by PDF libraries return different characters, without surfacing an user facing error. → most of pages need no OCR model. Text extraction costs milliseconds, performing OCR costs one or multiple inference steps and an accelerator. OCR on text only pages also lowers accuracy on tables. → We will discuss also model size differences and their impact on quality What a defect costs the agent: → We will discuss different ingestion paths and their affect on recall at 5. → What the impact of dropping 5 percent of letters is vs. removing 60% of every page are. → How expensive are duplicates? What about shuffled lines ?