News for hackers. Stuff that matters.

A comment on Launch HN: Midship (YC S24) – Turn PDFs and Images into usable data.

Re:Launch HN: Midship (YC S24) – Turn PDFs and Images into u... by serjester on Wednesday November 06, 2024 @04:09PM (#42069437)
Honest question but how do you see your business being affected as foundational models improve? While I have massive complaints about them, Gemini + structured outputs is working remarkably well for this internally and it's only getting better. It's also an order of magnitude cheaper than anything I've seen commercially.
[ Parent | on HN ]
Re:Launch HN: Midship (YC S24) – Turn PDFs and Images into u... by maxmaio on Wednesday November 06, 2024 @04:32PM (#42069753)
We're excited for foundational models to improve because we hope it will unlock a lot more use cases. Things like analysis after extraction, able to accurately extract extremely complex documents, etc!
[ Parent | on HN ]
Re:Launch HN: Midship (YC S24) – Turn PDFs and Images into u... by arjvik on Wednesday November 06, 2024 @06:32PM (#42071234)
Curious - have you compared Gemini against Anthropic and OpenAI’s offerings here? Am needing to do something similar for a one-off task and simply need to choose a model to use.
[ Parent | on HN ]
Re:Launch HN: Midship (YC S24) – Turn PDFs and Images into u... by serjester on Wednesday November 06, 2024 @07:54PM (#42072011)
Gemini is an awful developer experience but accuracy for OCR tasks is close to perfect. The pricing is also basically unbeatable - works out to 1k 10k pages per dollar depending on the model. OpenAI has subtle hallucinations and I haven’t profiled Anthropic.
[ Parent | on HN ]
Re:Launch HN: Midship (YC S24) – Turn PDFs and Images into u... by arjvik on Wednesday November 06, 2024 @09:49PM (#42072839)
Thanks!
[ Parent | on HN ]
Re:Launch HN: Midship (YC S24) – Turn PDFs and Images into u... by inapis on Thursday November 07, 2024 @02:44AM (#42074479)
If I may ask which model are you using? I have tried OCR'ing my bank statements in AI studio and the results have been less than optimal. Specifically it has a tendency to ignore certain instructions combined with screwing up the order.

Some pointers on what worked for you would be greatly appreciated.

[ Parent | on HN ]