PDF Tables with RowSpan and ColSpan not interpreted correctly

[nkidambi]

PDF Tables with RowSpan and ColSpan not interpreted correctly

[dayland]

@nkidambi , Thanks for the sample. We have added this to our backlog as an improvement to be made. Unfortunately it will not be in our 0.4-Delta or v1.0 Releases that are pending.

[ArpitaisAn0maly]

Describe the bug I have attached a PDF (publicly available) here. On Page 3 there is a table for VA pension benefits. For the question 'Can you tell me the full eligibility rules for receiving VA pension", the answer returned is incorrect. I have attached a screenshot of the issue.

To Reproduce Steps to reproduce the behavior:

1. Install the vNext-Dev version of the Information Assistant using bge embedding model
2. Upload the attached pdf file
3. Ask the question - "Can you tell me the full eligibility rules for receiving VA pension" and see the error as indicated in the attachment below.

Expected behavior A clear explanation of legibility conditions that should match the ALL the content in Page 3 of the attached document.

Screenshots

Alpha version details

• GitHub branch: vNext-Dev

Additional context summaryofvanationalguardandreserve.pdf

I have tested this particular use case and we are getting correct responses with GPT-4 in IA. GPT 3.5 as expected is not able to analyze tables to all details that is expected here. Given the variation in performance between GPT-3.5 and GPT-4 and the influence of chunking strategy on generation, it might be beneficial to explore chunking strategy for your particular use case. While IA works well with GPT-4 for your particular use case, we are seeing consistencies with GPT-3.5 with tables. You may have to tweak your top K to very limited to get what you need with GPT-4. Adding functionality to analyze tables with GPT 3.5 is in our backlog list. Thanks.

[nkidambi]

Thank you!