faiss vectore store and chatgpt - how do i use it best?

hello together!
i am experimenting with the “faiss vector store” in conjunction with chatgpt.
I am not quite clear yet

  • when and how the “agend prompter” accesses the store.
  • which size the documents for the store may have
  • and what is the best method or preprocessing to prepare documents for the store.

My status:
if I read a “fictitious term” with a short explanation as *txt into the store, then the agend prompter uses this info.
If the same information is attached to a large pdf document, which is preprocessed and divided into lines (about 400) with about 2000 words per line, then the agend prompter does not refer to it.

is there any info about the “faiss vector store” and how best to fill it with content?


What does your workflow look like?
faiss by default is an in memory vectorstore if not saved local. Do you have enough memory?

Hello Daniel,
this is my workflow and the table (csv) in which the split document is. this is also read in without problems. however, no good answer is given for the fictitious word “quademadelix” (located in the last line) in use case 2.

demo_vs_content.csv (1.1 MB)
Vector Store_demo.knwf (53.7 KB)


seems to work when I try it

When I try your long document I get a rate limit error as I am on GPT free.
What I am missing is the Textsplitting before embedding. Is this handled automatically?

i did the text spliting in a separate workflow, used this solution for it: https://forum.knime.com/uploads/short-url/6QRVjJvAB3wdBtVKb7FrL5SkE6n.knwf (input was a 280 page pdf document).
because a lot of memory is needed, i decided to use the external solution.

In my case i can read the long document (have paid account), but nothing comes out.

Another problem is that the text in the rows should not be too short to keep the context.

The question for me is how to preprocess the documents to use them for the vector store.
My goal: for example, a FAQ based on internal documents.



  • even if i reduce the big document (uc 2) to the last 30 rows, the system does not give me a correct answer out
  • if i concatenate the shortened document to use case 1 and uc 1 is on top as a row, it doesn’t work either

Just a short note, I am an advocate of LLMs and have similar ideas but you should be very careful in regards of “internal documents” an non private LLMs. Make sure that your company is really ok with this!

1 Like

Thanks for the hint, can never point it out enough. I use for experimentation currently only public documents :slight_smile:

1 Like

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.