PDF to text: inspect the text layer, then choose a parser

Extract a candidate text layer

pypdf extracts existing PDF text; it is not an OCR engine. Its documentation explains why PDF text extraction can be ambiguous. The presence of characters does not prove completeness or correctness, and a fixed character-count threshold cannot distinguish every scan from a digital page.

Use this extraction as a first pass. Compare representative pages with their rendered images, especially tables, formulas, unusual fonts and pages containing scanned regions.

Documentation-based example; no runtime measurement is claimed.

# pip install pypdf
from pypdf import PdfReader

reader = PdfReader("document.pdf")
for page_number, page in enumerate(reader.pages, start=1):
    text = page.extract_text() or ""
    print(f"--- Page {page_number} ---")
    print(text)
    if not text.strip():
        print("No extracted text: inspect this page and consider OCR.")

Try a structured conversion when layout matters

Docling converts to a document object and can export Markdown. Configure the chosen OCR and parsing pipeline for scans and mixed pages. A successful export is not a guarantee that every table or formula was preserved.

Documentation-based example; no runtime measurement is claimed.

# pip install docling
from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("document.pdf")
print(result.document.export_to_markdown())

MinerU offers another documented conversion path

The checked MinerU 4.0 interface supports stateless conversion with an explicit tier. Existing older installations should follow the migration documentation. The runtime and selected models affect results and resource use.

Documentation-based example; no runtime measurement is claimed.

# In an isolated environment, for the checked major-version API:
pip install "mineru>=4.0,<5"
mineru-kit parse document.pdf -o document.md --tier standard

Measure the output you will actually consume

For transcription, score CER/WER on declared ground truth. For structured pages, also score reading order, headings, table cells and formulas. OCRBench includes visual text understanding tasks, so its aggregate is not a universal plain-transcription error rate.

For search or RAG, inspect whether chunks keep headings, tables and page references together. Retain original pages and source locations so a downstream answer can be checked. Record package locks, model revisions, settings, failed pages and latency with any published evaluation.