Extraction confidence
How the app tells you what AI read correctly and what's worth reviewing, and why that isn't the same as a recipe's validation score.
When AI reads a document, it isn't always equally sure of what it understood. The app doesn't show you that certainty as a number: it translates it into warnings and lines to review, which point to exactly where to look.
Where you see it
- On each import's card (Upload new tab): when it finishes, the card says how many lines are worth reviewing ("2 lines to review" or "No lines to review") and, if there are any, the warnings after "warnings:" (for example, "warnings: Sheets to pick or create, Unit needs review"). See Import a photo or PDF.
- In History: each import shows its warnings below the status: "Sheets to pick or create", "Ambiguous unit", "Unit needs review", "Line without quantity"… See Import history.
- In the recipe editor: Open recipe opens the recipe with the validation panel expanded, and the flagged lines tell you what's wrong with them (no sheet, several possible sheets, assumed unit, no quantity).
The AI does work out internally how far it trusts its reading of the document type: it only rejects a file as "Not a recipe" when it's sure it's an invoice, a delivery note, or some other document. If it's unsure, it carries on and extracts the recipe, so a real recipe isn't blocked.
An import without warnings doesn't guarantee the recipe is correct: it only means AI didn't find anything doubtful. Always review amounts, units, and ingredients before you go into production.
Confidence isn't the validation score
They're two different things, and it pays not to mix them up:
- Confidence (the AI's) — how much the AI trusts that it read the document correctly. It's expressed through the warnings and the lines to review, at import time.
- Validation score (the recipe's) — a technical assessment of whether the recipe is well formulated (balances, consistency of amounts, etc.), which lives in the editor and in the formulation tools.
A recipe can be imported without a single warning—AI read it perfectly—and still need formulation adjustments. And the other way round: a reading with doubts doesn't mean the original recipe is wrong, only that the image was hard to read.
What to do based on the warnings
- No warnings — the reading was clean; even so, take a quick look in the editor.
- "Sheets to pick or create" — in the editor, pick the catalog sheet for each flagged line, or create a new one.
- "Ambiguous unit" or "Unit needs review" — confirm the units and amounts on the affected lines.
- "Line without quantity" — type in the amount that was missing from the original.
- "Not a recipe" (status) — you probably uploaded another type of document; you weren't charged anything. Upload the right file.
Import history
The inbox with all your photo and PDF imports: filter by status, review the warnings on each one, and open finished recipes.
Limits and best practices
What AI does well when importing recipes, what you shouldn't expect from it, and what you should always review before going into production.