Extract PDFs and scans with AI: check text, tables and numbers
A PDF that looks clear on screen may supply incomplete or disordered text to an AI tool. Before summarising an archive or filling a table, inspect what was actually extracted and preserve links to original pages.

Separate three operations
OCR recognises characters in an image. Layout analysis reconstructs blocks and tables. Structured extraction assigns values to fields. Success at one stage does not validate the next.
- Original preserved
- Values checked against the page
- Uncertainty stays visible
From document to verifiable data
Locate the first failing stage before switching models. Correctly recognised text can still be put in the wrong order or assigned to the wrong field.
Text
Compare a scanned page with one containing selectable text. Empty output does not establish that the original page was empty.
Structure
Read columns in their intended order. Match each amount to its heading and retain notes that qualify its meaning.
Fields
For a fictional invoice, distinguish net amount, tax and total. A missing value remains missing; it is not zero.
Verification
Retain a page or passage reference and check calculations. JSON formatting does not certify factual accuracy.
Which processing path fits your document?
Describe the file without sending it. This guide helps plan a trial; it does not predict software quality or process documents.
| File | Start with | Then check |
|---|---|---|
| Selectable text | Text and structure extraction. | Column order, headings and notes; a selectable PDF may still contain useful images. |
| Scan or image | OCR, followed by structure recovery. | Numbers, accents, negations and empty cells against the original page. |
| Mixed document | Page-by-page diagnosis, then suitable processing. | Coverage of every page and consistency between native and recognised text. |
For structured fields, add rules for types, units and missing values, then check values against the source page. For local processing, inspect OCR, vision models and plugins too.
Two approaches to examine: Docling · MarkItDown. MarkItDown targets text preparation for analysis; Markdown conversion does not guarantee faithful presentation. Read the official scope.
From document to usable data

AI-generated illustration 1. Identify the document
Select and copy a passage, then compare it with the screen. Distinguish text PDFs, scans and mixed documents. Locate rotated pages, columns, annotations and tables. Check rights and sensitivity before upload.

AI-generated illustration 2. Prepare pages
Work on a copy. Straighten pages and inspect contrast, noise and clipped areas. Tesseract documentation describes the effect of image quality and segmentation on recognition. Do not alter the image to match presumed text.

AI-generated illustration 3. Define fields
Write a simple schema: identifier, value, unit, page and supporting excerpt. Allow missing values and uncertainty notes. Distinguish zero, empty, unreadable and not applicable.

AI-generated illustration 4. Test difficult cases
Include a clear page, a multipage table, footnote and unclear value. Compare with human transcription. A polished summary must not hide incorrect table reading.

AI-generated illustration 5. Reconcile values
Check signs, decimal separators, units, dates and headers. Recalculate totals where available. Keep raw values before conversion. Do not infer missing data merely to make totals agree.

Editorial illustration 6. Validate before automation
Define critical fields requiring systematic review. Measure correction time per document and retain a control sample. Changes to engine, prompt or source format require another trial.
Choose conversion for the required output
Markdown may be enough for text analysis if essential information survives. MarkItDown offers this conversion; check installed components. For layout analysis and structured representation, examine Docling and its exports. Compare both on your files: documented functions do not constitute a fidelity ranking.
A reproducible table check
Prepare three fictional pages: a simple table, a table split across pages and a two-column page with a footnote. For each critical cell, record expected value, raw extracted value, heading, unit and page. Count incorrect values, omissions and wrong associations separately. A correct amount assigned to the wrong customer remains a blocking error.
Check conversion before summarisation
Keep three separate artefacts: original, converter output and model response. If an amount disappears, find the stage where it is lost. Changing a summary prompt cannot restore information that conversion never supplied. Also open the export in another tool and count correction time.
From scanned PDF to the output you need
For a searchable scan, try OCRmyPDF and check PDF search against the visible characters. To transcribe a region or obtain its coordinates, try Tesseract. To extract cells and their headers, compare a layout pipeline such as PaddleOCR or Docling on the same tables. For a scholarly bibliography in TEI XML, compare GROBID output with the PDF citations. Score words, structure and provenance separately.
Put the method to work

Practical case
Prepare five authorised pages with two tables, a rotated page and an unreadable value. Define ten fields without filling missing values.

Evidence to keep
Keep originals, raw text, normalised values, page numbers and human corrections.

Make the decision
Accept the batch only after reviewing every critical field and leaving unknown values explicitly unknown.
Detect correct values in the wrong place
Build three demonstration pages without personal data. Compare each output with the original before requesting a summary.

Two rows, two customers
- Input to prepare
- A table assigns customer A 120 euros and customer B 210 euros.
- Expected result
- Each amount remains attached to the correct row and customer.
- If the check fails
- Inspect cells and headings; a correct total alone cannot validate associations.

The footnote
- Input to prepare
- A fictional offer has an asterisk referring to an exclusion at the page bottom.
- Expected result
- Extraction preserves the note and allows it to be linked to the offer.
- If the check fails
- Inspect reading order and the link between marker and note text.

The page without a result
- Input to prepare
- A document contains a text page, a scanned page and a deliberately unreadable page.
- Expected result
- Processing distinguishes successful extraction, required OCR and unrecoverable content.
- If the check fails
- Reject overall success if a page silently disappears; expose limitations page by page.
Decision rule. Validate coverage and associations before summary quality. Retain original, raw extraction and correction decision separately.
Measure extraction

Fidelity
Exact values compared with the page before normalisation.

Traceability
Every field has a verifiable page and excerpt.

Structure
Cells remain linked to the correct headers.

Useful cost
Preparation and correction time per accepted document.
Explore document tools
These tools cover documents, local models and workflows. Check actual supported formats and quality on your own sample separately.
OCRmyPDF
searchable scanned PDF
OCRmyPDF contributors
Visit official siteTesseract OCR
optical character recognition
Tesseract contributors
Visit official sitePaddleOCR
document conversion
PaddlePaddle
Visit official siteDocling
document conversion
Docling project
Visit official siteMarkItDown
document conversion
Microsoft
Visit official siteGROBID
structured scholarly-document extraction
grobidOrg
Visit official siteHow is this selection produced?
These profiles are linked to the guide’s subject in our editorial catalogue. Inclusion is neither a ranking nor a performance measurement on your documents. Methodology.
Related tool families
Frequently asked questions
Does every PDF need OCR?
No. Some PDFs already contain text. First test extraction and reading order; OCR adds a stage that may introduce errors.
How should an unreadable cell be handled?
Leave it unknown, identify the page and request review. A plausible value is not a transcription.
Can important amounts be automated?
Define controls suited to their consequences, including critical-field review. A high overall score can hide a rare, costly error.
The references below expand on the concepts and checks discussed. Scenarios and trial frameworks remain editorial proposals; provider documentation describes its own product rather than an independent benchmark.
Source note
OCR and AI: check passages extracted from scanned PDFs
An image PDF does not necessarily contain faithful page text. OCR can confuse characters, columns, notes and tables. The U.S. National Archives note that OCR is often less reliable than manual transcription, particularly for handwriting or marked documents. Inspect a sample before requesting an AI summary.
Preserve the connection between extracted text and source image: page number, region and any corrected version. For important contractual or numerical information, check every decisive passage against the scan. Disclose uncertain reading instead of allowing AI to turn an OCR error into a confident claim.
Include difficult pages, not only a clear first page: tables, notes, columns and faint passages. Record the page and confusion type for each observed error. Decide which passages need manual correction before summarisation.
In extracted tables, check the relationship between label, value and unit. A correctly recognised number in the wrong column is false information. Keep corrected text separate from raw extraction and link corrections to the image for independent review.
- Input
- Check scan quality and reading order.
- Extraction
- Compare names, numbers and tables with the image.
- Output
- Summaries with page references and visible uncertainty.
Is searchable OCR text enough as evidence?
No. For important data, return to the page image and check that extraction has not moved or altered characters.
Fictional example — a two-column table
Extraction recognises every number but swaps two columns. The reader checks headings against the scan, corrects value associations and cites the page before requesting a summary.
Try the reasoning
OCR reads “1 500” as “1500” but assigns it to the wrong column. Which check should be added?
Reasoning: Preserve table position and headings. A correctly transcribed number can still be attributed to a different measure.
Primary source : U.S. National Archives — Citizen Archivist FAQ



