Llama Parse Accurate Mode Not Working Recently

if you could give a brief on the problem you are facing currently and share your job ID that would help @Sacha Bron to look into this 🙏
cc: @Sacha Bron

Accurate mode: 201a6d16-ec0a-4ba4-b6c7-219de06b6e78

Premium mode: 33cbb16f-3752-4340-80ad-3b098033ae9f

The thing is, accurate mode is struggling to execute simple parsing instructions... i feel like 2 weeks ago it was way better but i could be wrong

Hi @versa. We had a bug in our handling of the parsing instruction that we fixed, but unfortunately it broke your pipeline and you're not alone.
We're currently adding an option to make it work like it use to vs. the current behaviour.

Hi @Sacha Bron

Good to hear that you guys are working on it.

For now i'll stand with premium mode, which is absolutely amazing.

Thank you guys for the great work.

Thank you for giving us some feedback! Feel free to contact us again 😉

Hi @versa,
We just release a "fix" on Accurate mode. By default, your parsing instruction should work as it used to 2-3 weeks ago.
I added the is_formatting_instruction option which is True by default: https://docs.cloud.llamaindex.ai/llamaparse/features/parsing_options#is-formatting-instruction
If you need a parsing instruction like "translate the output to French" while keeping the clean markdown representation of the document, turn it off.
If you want to shape the output with a formating instruction like "only output a Json with this structure...", keep it on.

Amazing! Appreciate it

should "is_formatting_instruction" option be available on llama-parse 0.5.9 ? can't see it here

Yes, sorry it's only on the API and frontend yet.

I can probably release it in the next 24h

@versa It's released on v0.5.10: https://pypi.org/project/llama-parse
By the way, the is_formatting_instruction should be True by default so you may already have results similar to 3 weeks ago.

Great! Appreciate that