search

run-llama / llama_parse

Parse files for optimal RAG
https://www.llamaindex.ai
MIT License
3.03k stars 289 forks source link

"continuing this pattern" in output, instead of full data #478

Open bradmitchler opened 1 day ago

bradmitchler commented 1 day ago

I'm using Premium Mode to parse complex technical pdfs, with parsing instructions to break the table down by row with labels. It does it properly for the first couple rows of each table, but then stops and just says "continuing this pattern..." instead of continuing and outputting the full table data.

It actually has in the context a question from the LLM processing it "[Would you like me to continue with more detailed breakdowns of specific pressure points or flow rates for any particular model?]". I've tried editing the prompting it to always continue.

BinaryBrain commented 16 hours ago

Hi @bradmitchler, Can you please provide a JobID?

bradmitchler commented 15 hours ago

Here's an example: d4697f95-957d-415b-b192-5577fc286429

BinaryBrain commented 14 hours ago

Looking at your prompt, I think you may have more consistent results by setting is_formatting_instruction to false, as you don't really want to change the format of the output but simply give some directions.

bradmitchler commented 13 hours ago

I tried it - but then it doesn't break apart the data tables in to rows of text. I'm specifically trying to modify the tables to display in rows of text, like the output of that job I sent. It's doing it correctly, but not continuing for all cells / data points.

If I turn off formatting instruction, it leaves the data tables fully in tact. Thoughts?

BinaryBrain commented 13 hours ago

Yeah, LlamaParse's purpose isn't really to reformat data. We have this feature because it's a nice hack for some use cases. @hexapode I think this is a great use case for LlamaExtract