Best practices

dextract reads what is on the page and tells you which readings deserve a person. How many of them there are is mostly decided before the file reaches us, in how it was captured and what happened to it on the way.

Where dextract fits

dextract is not a black box that either works or does not. It reads every document and then tells you, document by document and field by field, which ones a person needs to see. That flag is the product: it is what lets you put reviewers where they change the outcome instead of on everything that arrives.

The split is yours, not ours: how much clears depends on how your documents arrive and where you set your own threshold. Everything below is about moving more of your volume onto the top branch.

needs_review

This is the field to build on. We do not reject documents for quality, because a rejection helps nobody: you still have the document and you still need what is on it. Instead every result carries a flag that says whether it should be seen by a person, and that flag is what makes the work plannable.

Signal	Where	What it means
needs_review	page	Look at this page. Most often it came in below 150 DPI, where a value can land on the row above its own however cleanly it printed.
needs_review	field	Look at this value. It is uncertain on its own terms, whatever the rest of the page looks like. On a checkbox group it means the marks were ambiguous.
confidence	field	How clearly the characters read, 0 to 1. We emit it; what you set is the line you draw through it.

Signal

Where

What it means

needs_review

page

Look at this page. Most often it came in below 150 DPI, where a value can land on the row above its own however cleanly it printed.

needs_review

field

Look at this value. It is uncertain on its own terms, whatever the rest of the page looks like. On a checkbox group it means the marks were ambiguous.

confidence

field

How clearly the characters read, 0 to 1. We emit it; what you set is the line you draw through it.

Confidence and needs_review answer different questions. Confidence says how legible the text was. needs_review says whether the reading is worth checking. Clean printing read off the wrong line scores high on the first and is exactly what the second exists to catch, so a pipeline that branches on confidence alone will let that one through.

The two levels are independent. A flagged page does not flag its fields, and a flagged field does not raise the page, so check both: the page flag decides whether the document gets opened, the field flags decide what the reviewer looks at once it is.

Nothing is hidden behind the flag. A flagged document still comes back fully extracted, with every value and every box, so review means confirming a reading rather than starting again.

Setting your threshold

Route on the field, not the document. One uncertain box on an otherwise clean claim should send one field to a person, not the whole file.

Start by reviewing everything for a week and record which fields your reviewers actually changed. That tells you where your own threshold belongs, which is almost never the same for a total charge as for a middle initial. It also tells you what the flag is worth on your documents, which is a number you should have before you rely on it.

Every value carries the box it was read from, so a reviewer can confirm it against the document rather than re-keying it. That is the difference between checking a result and redoing it.

Resolution

Send at least 150 DPI. On a letter-size page that is roughly 1275 pixels wide, and 200 DPI or 1700 pixels is better.

Resolution matters more than it sounds because dense forms print their lines close together. A claim form at 150 DPI puts about 15 pixels between one line of a box and the next. At 80 DPI that gap is 7 pixels, and the value on one line becomes hard to tell from the caption above it. We still read the document, but the chance of a value landing on the wrong field goes up sharply.

We measure the resolution of every page against the form we matched it to. When it falls below 150 DPI we mark the page for review rather than refusing it, so you get the extraction and a clear signal that it deserves a look.

File formats

We accept PDF, PNG, JPEG, TIFF, BMP and WebP.

Send the original. If your document is already in one of those formats, upload it as it is. Converting it, re-saving it, or resizing it to fit an email or a form field costs detail that cannot be recovered, and it is the most common reason a document that should read perfectly comes back needing review.

A native PDF is the best input there is, because its page is rendered at full resolution rather than resampled from an image. If you have both a PDF and a scan of the same document, send the PDF.

Avoid re-encoding a JPEG that is already a JPEG. Each save discards a little more, and the artefacts land exactly on the thin printed rules that tell one field from another.

Scanning and photos

Scan flat at 200 to 300 DPI in colour or greyscale. Bitonal or "black and white" scanning at low resolution removes the faint print that many forms use for their own field guides.

Photographs of documents work, with care. Fill the frame with the page, keep the camera parallel to it rather than at an angle, and use even light. We correct for a moderate skew, but a photograph taken at a steep angle stretches one side of the form relative to the other in a way that shifts values between neighbouring boxes.

Keep the whole page in frame, including the margins. Some forms carry information at the very edge, and a crop that clips it also removes landmarks used to line the page up.