Yeah, I wrote about those limitations in my post https://
simonwillison.net/2024/Mar/30/oc
r-pdfs-images/
… It's not state-of-the-art, but it's aiming to be the most convenient possible (free) option for people who don't have anything better
@simonw
-
Free OCR tool for PDFs and images with practical limitations
By
–
-
LLM Continuation Trick: Requesting More Output Within Context
By
–
An interesting trick that does work is you can send a prompt requesting "more" and have the LLM pick up again where it stopped That requires round-tripping the work it has done so far, but with a long enough context window (and a will to spend the money) that's quite feasible
-
LLM Output Token Limits: GPT-4, Claude 3, Gemini Pro
By
–
What are the LLMs with the most output tokens these days? GPT-4 and Claude 3 are both 4096. Gemini Pro 1.5 is 8192 This really matters for structured data extraction: even with 1m of input tokens you can't scrape a big webpage into a CSV file if you run out of output tokens
-
Claude 3 Opus Vision Models Refuse PII in PDF Analysis
By
–
I have been doing a bunch of experiments with the latest vision models and PDFs recently, they're pretty great! Claude 3 Opis did point-blank refuse to answer questions about one that included personally identifiable information though
-
Effective Chunking Strategies for GPT-3.5 Implementation
By
–
Have you found a chunking strategy that works well powered by gpt-3.5? I am hoping to add chunking to LLM at some point but I'm not sure what strategies I should include
-
Browser-based packages loadable from CDN networks
By
–
Do you know anything like that which runs directly in the browser via a package you can load from a CDN?
-
Expanding Tesseract.js Language Support in OCR Tool
By
–
Tesseract.js works with other languages, but I've not exposed that in my tool yet – I should do that!
-
textract-cli: Lightweight AWS Textract Command Line Wrapper
By
–
My other OCR project from yesterday: textract-cli, a tiny CLI wrapper around AWS's amazing but so-hard-to-use Textract API https://
github.com/simonw/textrac
t-cli
… Assuming you have AWS credentials configured: pipx install textract-cli
textract-cli image.jpeg > output.txt <5MB JPEG/PNGs only -
Privacy-First JavaScript Project: Local Processing No Servers
By
–
That's more sophisticated than I want to go with this particular project – the goal here is privacy-first, no servers, document stays on your computer and all the work happens in JavaScript running in your browser
-
AWS Textract OCR: Best Tool But Difficult User Access
By
–
Totally agree, AWS Textract is still the best OCR tool I've used It's a shame it's so difficult to put it in people's hands though! Just signing up for AWS billing is enough to put most people off