Convert, then refuse bad outlines
Use MarkItDown Online to convert PDF, DOCX, PPTX, and related formats, then inspect the heading outline in preview. Reject files where hierarchy is flat, inverted, or padded with repeating chrome. Re-convert or clean the source before the file enters an agent loop.
Protect lists and tables for tools
Agents that cite rows or call tools need list items and table structure to stay grouped. Prefer a slightly imperfect Markdown table over merged cell blobs. If bullets collapsed into one paragraph, fix that before ingestion.
Batch with timing plus outline review
For multi-file uploads, convert as a batch, skim durations and outlines together, then download the ZIP. Two minutes of preview review prevents hours of bad embeddings. Keep the gate boring and consistent: outline first, then download.