Issue with LlamaParse: Hallucinations on Bank Statements #303

Closed
opened 2026-02-16 00:17:24 -05:00 by yindo · 3 comments
Owner

Originally created by @bsuttonvyde on GitHub (Oct 18, 2024).

Originally assigned to: @BinaryBrain on GitHub.

Note: we're aware of some missing content in the output and layout issues on tables. Please refrain from opening new issues on this topic unless if you think it's different from what has already been reported.

Describe the bug
I am parsing 4-10 page bank statements with "Accurate" mode and it is creating transactions that don't exist on the original bank statement document.

Files
I'd prefer not to upload the document as it has personally identifiable information.

Job ID
8e57c117-1263-4eed-9131-3a50c05c12ee

Options
I used "Accurate" mode with parsing instructions.

Originally created by @bsuttonvyde on GitHub (Oct 18, 2024). Originally assigned to: @BinaryBrain on GitHub. _Note: we're aware of some missing content in the output and layout issues on tables. Please refrain from opening new issues on this topic unless if you think it's different from what has already been reported._ **Describe the bug** I am parsing 4-10 page bank statements with "Accurate" mode and it is creating transactions that don't exist on the original bank statement document. **Files** I'd prefer not to upload the document as it has personally identifiable information. **Job ID** 8e57c117-1263-4eed-9131-3a50c05c12ee **Options** I used "Accurate" mode with parsing instructions.
yindo added the bug label 2026-02-16 00:17:24 -05:00
yindo closed this issue 2026-02-16 00:17:24 -05:00
Author
Owner

@bsuttonvyde commented on GitHub (Oct 18, 2024):

Does the use of parsing instructions increase the likelihood of hallucinations?

@bsuttonvyde commented on GitHub (Oct 18, 2024): Does the use of parsing instructions increase the likelihood of hallucinations?
Author
Owner

@BinaryBrain commented on GitHub (Oct 18, 2024):

It does because we give a high priority to your prompt compare to our default formatting. We've just released a flag that could help you: https://docs.cloud.llamaindex.ai/llamaparse/features/parsing_options#is-formatting-instruction

It's enabled by default but if you disable it, you should have a cleaner markdown with less to none hallucinations.
Also, if your prompt is too long, it can actually decrease the quality of the output because of the limited attention of the model.

@BinaryBrain commented on GitHub (Oct 18, 2024): It does because we give a high priority to your prompt compare to our default formatting. We've just released a flag that could help you: https://docs.cloud.llamaindex.ai/llamaparse/features/parsing_options#is-formatting-instruction It's enabled by default but if you disable it, you should have a cleaner markdown with less to none hallucinations. Also, if your prompt is too long, it can actually decrease the quality of the output because of the limited attention of the model.
Author
Owner

@bsuttonvyde commented on GitHub (Oct 18, 2024):

Thank you for that information!

@bsuttonvyde commented on GitHub (Oct 18, 2024): Thank you for that information!
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#303