TinkerSupplier invoice: Vision at 3,200 px. Plan total: 11 of 11 fields, 14,256 tokens.
Chapter 7 of 7
Workbench: pick a method and a size
Larkspur now reads three kinds of document every day: invoices, packing slips and the warehouse’s notes. Each costs tokens, and each can be read by OCR or by vision, at any size.
This chapter: Which method and size should each kind of document get?
You will learn
- How to choose a method and a size for each kind of document
- Why the cheapest setup that gets every field right wins
- Why the answer is different for each kind
Words you’ll meet (3)
- Plan
- A method and a size for each kind of document.
- Cost
- Here, the tokens the model reads: the image’s tokens for vision, the text’s tokens for OCR.
- Workbench
- The chapter’s free-play bench: change the plan and see every result.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 2
The proposal on the table: vision at the largest size for everything, to be safe. The bench reads the recorded reading for your plan, document by document, and grades every field live. The best plan gets the most fields right (11 is the most any plan gets) for the fewest tokens. Pick a method and a size for each kind of document.
11 of 11 fields right · 14,256 tokensBEST IS 11 FOR 3,076
| Document | Fields right | Tokens |
|---|---|---|
| Supplier invoice, Vision at 3,200 pxEvery field right, the same as at 1,600 px: the image was scaled down to the model’s limit first. | 5 of 5 | 4,758 |
| Packing slip, Vision at 3,200 pxThe model reads the table as a table: each number stays with its row and column. | 3 of 3 | 4,758 |
| Handwritten note, Vision at 3,200 pxThe model reads the handwriting right at this size, though handwriting stays a known weak spot. | 3 of 3 | 4,740 |
The readings were written for this site to illustrate the cited findings; they were not measured on a real model. Image token counts follow Anthropic’s published rule, computed live; OCR text is counted with this site’s tokenizer.