PDF

Very Early Stage

This code is in a very early stage.

Quickstart

# Load your pdf file
with open('apple_10k_2024.pdf','rb') as f:
    content = f.read()

# Convert to dictionary
dct = pdf2dict(content,mapping_dict=None)

Benchmarks

• About 200 pages per second single threaded.

multithreading

pdf2dict can't be run multithreaded due to the limitations of pypdfium2

Compatibility

Requires pdfs with underlying text structure so no scans yet.

convert_scan_to_instructions would be fairly straightforward to implement. Font-size can be inferred from bounding boxes, as can line alignment. Rotation probably won't be an issue for decent scans like the ones submitted to the SEC.

The issue is performance.

The point of doc2dict is mostly that it's fast. Local OCR such as pytesseract would put a hard cap of 10 pages per second.

This is too slow to be useful for my use-case. Here's a benchmark.

Convert all 2024 Annual Report to Shareholders to dict form 2000 a year * mean 50 pages / 200 pages per second = 500 seconds = ~ 10 minutes. (PDF Parser)

Where as a scan parser would take at least 200 minutes ~ 3 hours.

I think the solution will be to write a scan parser that takes input of bounding boxes/ minimum features required as input. Users can then use their preferred OCR - e.g. local, Google, AWS, etc for the slow part.

TODO

think about tables get center get other old attributes like indent

Issues

• Adobe PDF encodings return weird characters.