Tesseract OCR: Getting Text Out of Scans, Locally

What it is, and what it is not

Tesseract is an optical character recognition engine with a long history — decades of development, now open source under Apache 2.0. Since version 4 it recognises text lines with a recurrent neural network rather than character by character, which lifted its accuracy on ordinary documents substantially.

What it is not: a document understanding system. It returns text, optionally with positions. It will not tell you that a block is a heading, will not rebuild a table into rows and columns, and does not understand what it reads. For structured documents, OCR is one brick among several — the layout work belongs to a parser like Docling.

Language data: the first thing people get wrong

Tesseract reads English by default. Run it on a French, German or Spanish document without installing that language's data file and you get lost accents and near-miss words — from which many people conclude the engine is bad. Install the language pack and request it explicitly.

Situation	What to do
Non-English document	Install that language's data and pass it at runtime
Mixed-language document	Pass several languages together; accuracy drops slightly, coverage improves
Numbers-only field	Restrict the allowed character set to digits
Unusual script	Check a language pack exists before planning the project around it

Image quality does the heavy lifting

This is the lesson everyone learns late: on a mediocre scan no flag rescues the output, while ten lines of preprocessing turn unreadable text into clean text.

• Resolution — around 300 dpi is the target. Below 200 accuracy collapses; far above it you pay time for nothing.
• Deskew — a page tilted two degrees costs more accuracy than any engine tuning recovers.
• Contrast and binarisation — light grey text on a grey background should become black on white before it reaches the engine.
• Cropping — removing scanner borders and stray margins prevents phantom characters.

Page segmentation: the setting that returns empty strings

Tesseract tries to work out the structure of what it is looking at before reading it. The default assumes a full page of text. Hand it a single line — a label, a receipt line, a narrow column — and it may return nothing at all, which reads like a failure and is a misconfiguration.

Telling it what the image contains (a block of text, a single line, a single word) is often the difference between an empty result and a perfect one. It is the first thing to change when a crop that clearly contains text produces nothing.

Tesseract or a vision model

Criterion	Tesseract	Local vision model
Resources	CPU, tens of megabytes	GPU preferred, several gigabytes
Speed	Very fast	Much slower
Clean single-column text	Excellent	Equivalent, more expensive
Complex layout, tables	Weak	Much better
Understands content	No	Can answer questions about the page
Invents values	Never — it errs, it does not fabricate	Yes, plausibly formatted and wrong

That last row decides most document-control use cases. Tesseract produces visible garbage when it fails; a vision model can produce a clean, confident, incorrect figure. On accounting or administrative records that is a difference in kind, not degree. For layout-heavy documents where you accept that risk, PaddleOCR and vision models are the stronger tools.

Where it belongs in a local stack

In a local document pipeline, Tesseract appears at exactly one point: immediately before indexing, on files with no text layer. The test to automate is simple — if extracting text from a PDF yields fewer than a few dozen characters per page, it is a scan and goes to OCR. Without that check, whole documents enter the index empty and nobody notices until a user asks why the assistant has never heard of a contract that is right there. The rest of the chain is the usual one, ending in a local model answering from retrieved passages — see what RAG is.

Verdict

Tesseract remains the sensible default for the OCR step: free, offline, fast, and predictable in how it fails. Spend your effort on the images rather than the flags, install the right language data, and tell it what shape the page is. Reach for a vision-based reader only when the layout genuinely defeats it — and never on documents where an invented number would go unnoticed.