The reader self creates data out of thin air? #281

Open
opened 2026-02-16 00:17:21 -05:00 by yindo · 2 comments
Owner

Originally created by @kinshukkaura on GitHub (Sep 29, 2024).

Originally assigned to: @hexapode on GitHub.

Describe the bug
The reader creates data out of thin air for multiple pages. It creates tables with information that is not available anywhere in the pdf page.
.

Files
ppfas-mf-factsheet-for-August-2024.pdf

Job ID
7f63cc55-1a75-450d-aea0-3a6aa3c648ba

Screenshots
image
image

Client:

  • Frontend (cloud.llamaindex.ai)
  • Python Library
  • API

Options
Using the accurate method, with all other fields default/empty.

Additional context
Add any additional context about the problem here.

Originally created by @kinshukkaura on GitHub (Sep 29, 2024). Originally assigned to: @hexapode on GitHub. **Describe the bug** The reader creates data out of thin air for multiple pages. It creates tables with information that is not available anywhere in the pdf page. . **Files** [ppfas-mf-factsheet-for-August-2024.pdf](https://github.com/user-attachments/files/17180776/ppfas-mf-factsheet-for-August-2024.pdf) **Job ID** 7f63cc55-1a75-450d-aea0-3a6aa3c648ba **Screenshots** ![image](https://github.com/user-attachments/assets/f4954b58-19fb-4086-89d6-b89d608427de) ![image](https://github.com/user-attachments/assets/8478e749-9c30-42fd-b752-d2a8bec95110) **Client:** - Frontend (cloud.llamaindex.ai) - Python Library - API **Options** Using the accurate method, with all other fields default/empty. **Additional context** Add any additional context about the problem here.
yindo added the bug label 2026-02-16 00:17:21 -05:00
Author
Owner

@hexapode commented on GitHub (Sep 29, 2024):

Had a look at your job and you used our default mode (Accurate). This document work well with premium mode (see attached markdown, with the exception of a miss-classified chart as an image).
ppfas-mf-factsheet-for-August-2024.pdf.md

However the premium mode is more expensive as more compute is involved. Alternatively you can try to use our fast mode that will layout the text in an understandable way for LLM (but not extract the tables)

@hexapode commented on GitHub (Sep 29, 2024): Had a look at your job and you used our default mode (Accurate). This document work well with premium mode (see attached markdown, with the exception of a miss-classified chart as an image). [ppfas-mf-factsheet-for-August-2024.pdf.md](https://github.com/user-attachments/files/17182559/ppfas-mf-factsheet-for-August-2024.pdf.md) However the premium mode is more expensive as more compute is involved. Alternatively you can try to use our fast mode that will layout the text in an understandable way for LLM (but not extract the tables)
Author
Owner

@kinshukkaura commented on GitHub (Sep 29, 2024):

Thanks. I believe I would have to use the premium mode.
is there any reason why the model hallucinates in the default mode (Accurate) and not in other modes? Could playing with the parsing instructions (prompt) help in any way?

@kinshukkaura commented on GitHub (Sep 29, 2024): Thanks. I believe I would have to use the premium mode. is there any reason why the model hallucinates in the default mode (Accurate) and not in other modes? Could playing with the parsing instructions (prompt) help in any way?
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: run-llama/llama_cloud_services#281