Use text from images and screenshots carefully

Key takeaways

• Docsi reads PNG, JPEG and WebP locally with text recognition.
• Sharp screenshots are usually easier to read than photos of a screen.
• Photograph documents straight, evenly lit and without flash.
• Tables arrive as lines of text and need a row-by-row check.
• Check digits, names and amounts character by character against the image.

Provide a readable image

The file picker supports PNG, JPEG and WebP. Use a sharp image with all relevant lines. An original file is often better than a reduced screenshot copied from a message.

Docsi can recognise an image source’s text locally. That does not establish accurate interpretation of handwriting, charts, symbols or tables. For tiny text, tighter readable crops can help, provided context survives.

Separate recognition from interpretation

1. Add the image and name what you need from it.
2. First request the important visible text passages.
3. Compare names, numbers and special characters with the image.
4. Only then request explanation or structured output.

Leave absent structure unresolved

A table edge may clip important columns. An error screenshot does not automatically establish its cause. Identify visible evidence and, when needed, supply further images or the original file.

Turn image text into a source step by step

First inspect legibility without enlargement that merely creates bigger pixels. A new sharper capture is often more useful than enlarging a tiny image. Keep related lines together and name their order if several images are needed. For tables, capture headers and units with values. An amount without currency or a date without a field label cannot be interpreted reliably.

When correcting recognised text, state the correction explicitly. Request that confirmed reading and check it in the result. A screenshot is not an editable original: page structure, hidden comments or formulas cannot automatically be recovered. Use the actual source file for those questions. The image remains evidence of what was visible.

Which images Docsi accepts

Docsi accepts PNG, JPEG and WebP as image sources, the usual formats for screenshots and phone photos. He scales large photos to a compact copy of at most 2,400 pixels per side; the original stays unchanged. The text is recognised locally with Tesseract for German and English.

If the chosen AI model can process images, it also receives the image in an AI task. That helps with understanding layouts but does not replace checking the recognised text. Convert iPhone photos in HEIC format to JPEG beforehand, for example via “Export” in the Photos app.

Taking good screenshots

On Windows, open the Snipping Tool with Windows key + Shift + S and select an area; on a Mac, Cmd + Shift + 4 captures a selection. Crop to the relevant part instead of capturing the whole screen; that keeps the text large enough after scaling.

Zoom the view to 125 to 150 percent before capturing if the text is small. Avoid semi-transparent windows, mouse pointers over the text and dark themes with grey text. A screenshot with clear contrast almost always gives better results than a photo of the screen.

Photographing documents

Lay the sheet flat on a dark, plain surface and hold the phone parallel above it. Use daylight or even room light and avoid the flash, which washes out glossy paper. The sheet should largely fill the frame without cutting off any edges.

For multi-page material, photograph each page individually and number the images consecutively. For longer documents, a scan to PDF is more practical: Docsi then reads the pages with page numbers and you can check them one by one.

Taking tables over from images

Text recognition delivers lines, not cells. A table in an image therefore arrives as a sequence of text lines in which columns can shift. Ask Docsi to build a table with explicitly named columns from it, and compare every row with the image.

With many rows or important figures, it is often quicker to get the original data – for example as an Excel or CSV download – or to type the values yourself. An image is a good source for text but an uncertain one for columns of figures.

Checking recognised text deliberately

Typical errors in image text are confused characters (0 and O, 1 and l, rn and m), missing commas in amounts and merged words at line ends. Read figures, names and dates character by character; running text tolerates small errors better.

The recognition confidence shown is an engine estimate. Even with high values, individual digits can be wrong. So only accept text you have checked as the basis for source markers and analyses.

A worked scenario

An example: after a lecture you photograph the whiteboard with a formula derivation in three steps. You take three photos instead of one, each straight from the front and without reflections. Docsi recognises the printed part well, but the handwritten formulas only patchily.

You type the formulas yourself and add them as text; the photos serve as evidence of the order. Then you ask Docsi to explain the derivation step by step and refer to the photos. That way you combine the strength of text recognition with print and your own checking for handwriting.

Checklist before sharing

Text from images often looks right but contains individual wrong characters. This check focuses on the places where errors do the most harm.

1. Numbers, amounts and dates have been checked character by character.
2. Names and technical terms match the image.
3. Table rows have been compared with the image one by one.
4. Illegible passages are marked instead of guessed.

Key terms explained

Screenshot

An image of the screen content or part of it.

Cropping

Removing unnecessary image margins so the relevant text stays larger.

Vision model

An AI model that can process images as well as text.

Contrast

The brightness difference between text and background; high contrast improves recognition.

HEIC

The default photo format of many iPhones; convert to JPEG before adding.