Smash Notes

Back to The Missing Infrastructure for AI Agents: Documents | Reducto’s Adit Abraham, with Adit Abraham & Jennifer Li.

How Does Reducto Read a Messy Document?

Adit Abraham breaks document reading into a few epochs, ending with Reducto R1, one unified model that reads a page from end to end. Bounding boxes stay in place, the result stays consistent, and the model captures everything on the page without the accuracy and latency trade-offs older setups had to accept.

PDF software started as a way to print. People in the 1990s wrote drivers so every character would appear on the page. OCR companies later templated a steady form such as a W-2, pinning the pixels where a name sits and cleaning the text afterward. Templates were enough for a lot of consistent pipelines. In 2023, nobody thought about complexity before uploading a document to ChatGPT. The file was simply expected to work. Visual language models reason over the contents, and a document is an image: a 9 by 11 grid. Two columns have to be read in order. A table is two dimensions of data, and the rows, columns, and merged cells all have to survive.

Reducto began as a two-part problem. Traditional computer vision preserved structure, and it was strong at determinism and traceability, because customers need to know where an answer came from. Visual language models ran as a correction loop where accuracy and latency still had a cost. The expectation and the stronger tools arrived together around when Reducto was founded. R1 now does the whole job in one model.

Grokipedia·Reducto·ChatGPT

Transcript

The Missing Infrastructure for AI Agents: Documents | Reducto’s Adit Abraham ·

From the transcript

I think there's still quite a few problems to be solved. Break down the world of Redacto's documents for us. Yeah, I always joke that I've literally met people that have worked on PDF processing longer than I've been alive. There were people in the 1990s that were working on drivers to print PDFs, which was the original purpose of the format, right? Like how do you faithfully represent all the characters on the page? You can kind of break this down into a few epochs.

Like first it was a file processing problem. Then you started having these like OCR based IDP companies. And typically what you would have done is you have something like a W-2. You template out what a W2 looks like, like you have a pixel level understanding of the name appears here and you post-process that text. And that was fine for a lot of consistent pipelines.

See in transcript

a16z

More from this episode

  1. Why Do Document Mistakes Compound Across a Dozen Files?Adit Abraham says a loan packet is more than reading the page.
  2. Why Did Coding Tools Reach a Billion in Revenue First?Adit Abraham says coding tools reached a billion dollars in annual revenue while professional-services AI has not, even though healthcare, insurance, legal…
  3. Why Won't the Big Model Labs Solve Document Parsing?Jennifer Li asks why the big model labs, already focused on knowledge work and on vision, will not simply solve document parsing.
  4. Do Frontier Models Silently Drop Facts From Long Documents?Adit Abraham says frontier models were more precise than many specialized document parsers on a long-extraction benchmark: when they pulled a fact, it was…
  5. What Should You Check First in a Document Benchmark?Adit Abraham says the first thing to check is whether the distribution of data matches the work you want done.
  6. Can Agents Do Healthcare Work Without Reliable Document Parsing?Jennifer Li asks how accuracy changes whether customers will trust an agent with their documents.

People

More answers like this

Related episodes

Plays from this note through the rest of the episode.

Weekly highlights

One letter a week.

The notes worth hearing, with the minute they were said.