Most teams optimize prompts and models but ignore the biggest source of waste: messy document input.
Where the tokens go
- Raw HTML can be 3–5× larger than the equivalent Markdown.
- PDF text dumps include repeated headers, footers, and noise.
- Unstructured input forces the model to work harder, increasing output tokens too.
The fix
Convert documents to clean Markdown before they hit your model. You’ll cut input tokens, reduce hallucinations, and speed up responses — for free.