Start with pages → tokens
Digital PDFs often land around a few hundred tokens per page after text extraction; dense contracts run higher, slide decks lower. Treat tokens-per-page as a planning knob you calibrate on a sample, not a universal constant.
Chunk overlap inflates the embed bill
RAG pipelines rarely embed the corpus once as a single string. Overlapping chunks improve retrieval but send some tokens more than once. A 10–20% overlap is a common starting point for cost models.
Extract passes dominate when OCR is involved
Scanned pages and messy layouts often need an LLM cleanup or structured extract. That is usually far more expensive than embedding. Prefer batch APIs for offline ingestion, and only run extract on the fraction of pages that need it.
FAQ
Does this include vector database storage?
No—CentsPerToken prices model API usage (embeddings and optional LLM extract). Hosting and vector DB fees are separate.