Skip to content

vera-ingest-pymupdf

Default PyMuPDF ingest pipeline plugin for VERA. Install it beside vera-ingest to register the pymupdf provider under the vera.ingest_pipelines entry-point group.

vera-cli and vera-app depend on this package so PDF conversion works out of the box.

Install

python -m pip install "vera-ingest-pymupdf>=0.3.0"

From a repository checkout with uv, the workspace installs it by default:

uv sync

The package depends on PyMuPDF and pdfplumber. English OCR works offline with the bundled eng.traineddata. Other Tesseract languages can be installed manually via TESSDATA_PREFIX or fetched on demand with --ocr-allow-download / vera ocr-languages download.

Usage

vera convert "manual.pdf" "manual.vera"
vera convert "manual.pdf" "manual.vera" --parser pymupdf

from vera_ingest import convert

convert("manual.pdf", "manual.vera", parser="pymupdf")

Behavior

  • Parses PDFs with PyMuPDF; extracts bordered tables with pdfplumber.
  • Selectively OCRs image-dominant low-text pages (ocr_mode=auto), with force for full-page OCR and off to disable OCR. Auto mode still OCRs a large-image page whose only native text is sparse (fewer than 200 alphanumeric characters: headers, Bates stamps, letterhead). Blank pages and native-text pages without a large image skip OCR.
  • Owns typed defaults: chunk_size=500 whitespace-split words, overlap=75 words, ocr_mode=auto, ocr_language=eng, ocr_dpi=300.
  • Sliding-window chunking with heading detection produces searchable text and figure metadata. Size and overlap count str.split() words, not characters or LLM subword tokens. Runtime overlap is clamped to chunk_size - 1.
  • OCR is designed for scanned prose and does not reconstruct complex scanned forms or tables.

Desktop app

Packaged and source-run desktop conversions use this pipeline by default. Convert controls are schema-driven from the pipeline descriptor (describe_ingest_pipelines / PipelineConfigForm). The sidecar build bundles the English tessdata directory from this package and copies package metadata for entry-point discovery; importing the package also calls ensure_registered() so PyInstaller freezes still resolve pymupdf when dist-info is missing.

See also