Pulpie aims to clean web pages at much lower cost
Feyn’s Pulpie is a group of models that takes raw HTML and removes repeated page clutter such as ads, footers, and sidebars. It returns only the main content as HTML or Markdown. Feyn says Pulpie reaches a similar quality to Dripper, a leading current extractor, while costing about 20 times less.
Cleaning 1 billion web pages is presented as costing $7,900 with Pulpie versus $159,000 with Dripper. The claimed savings come from the model design. Many leading extractors use a decoder approach, which creates output one token at a time and reads the full model from memory at each step.
Pulpie uses an encoder approach, which reads the full input HTML once and labels each block as either content or page clutter. This makes Pulpie rely more on computing power, while decoder tools rely more on memory speed. Cheaper GPUs often have relatively strong computing power compared with , so Feyn says Pulpie can run efficiently on lower-cost hardware.
Key points
- Pulpie removes ads, footers, sidebars, and other clutter from raw HTML.
- It returns the main page content as HTML or Markdown.
- Feyn compares the cost for 1 billion pages as $7,900 for Pulpie versus $159,000 for Dripper.
- Pulpie uses an encoder design that labels page blocks instead of generating text piece by piece.
- The cost advantage is tied to running well on cheaper GPUs.