Convert, then audit the outline
Use MarkItDown Online to convert PDF, DOCX, PPTX, and related formats to Markdown, then immediately check H1–H3 order. If every section is an H1, or body text was promoted into titles, fix or re-convert before the file enters an agent loop.
Keep lists and tables machine-readable
Agents that call tools or answer from RAG chunks need list items and table rows to stay grouped. Prefer a slightly imperfect Markdown table over a blob of cells. Reject outputs where bullets collapsed into a single paragraph.
Batch with a quality gate
For multi-file uploads, convert as a batch, skim heading outlines in preview, then download the ZIP. The quality gate is human and cheap: two minutes of outline review prevents hours of bad embeddings. Typical conversions still finish in about 2–5 seconds per common document.