INFO NO FOUND with a big document

[be-swarm]

Hi,
I do the following search "Retrieve the URL of the guide produced by ADEME in the audit document."
When I do this on the large document I systematically get an INFO NOT FOUND. When I do the same search on the small document, the result is ok.
Who can explain to me why?
Are there any settings to configure for this to work?
THANKS.
large document AUDIT.pdf
small document:
audit-light-ademe.pdf

[kbeaugrand]

Hi,

Your question is quite vague, means that the chunks retrieved might not contains the correct information
For example: Where can I find the ADEME guide that provides advice on managing interfaces between renovation work phases? might better work.

Otherwise, the solution could be during the ingestion phase. The document you provided is quite complex and have a dedicated structure that is not directly understandable by LLM. I used and prefer to work in custom ingestion pipeline to handle this kind of documents.

On another hand , you might try to test KernelMemory.StructRAG that might improve you search query.

BTW, your document contains personal data, be sure to remove it please.

[be-swarm]

Hi,
Thanks for your response.
Your example seems to give good results.
I'll look at KernelMemory.StructRAG
Your remark is relevant about personal data but the attached document only contains test data which is not real.

[JoelAuren]

@kbeaugrand I have used StructRAG and I like the results it gives, but is it normal for the search with StructRAG to be much slower than the normal kernel memory mode?

[kbeaugrand]

Hi,

Nice to see that StructRAG works and gives answers.
Yes right, since we have multi step and reasoning into StructRAG to take much times (about x2) but more important, my experiments told me that it increase the cost (number of tokens used) by x5.

So take care of that and where StructRAG is important or not.

[JoelAuren]

Thank you for notifying me about the cost increase (x5). I will enable a checkbox to activate or deactivate the StructRAG functionality only in cases where I need a more elaborate response. I also want to mention that I currently use the gtp4omini model, which is not very expensive.

[kbeaugrand]

yes the cost is important to have in mind, but since it's not so much expensive with this model, the overall cost might not be also huge, depending on the usage in fact.