@mgirdwood I have this example where I load several PDFs from the author Stefan Zweig into a Vector store and make them accessible. You might have to experiment with the meta data of the documents and also with the ‘window’ you give the model.
There is a component that tries to do a more complex preparation of the data with giving blocks of text (since a novel might benefit from a few sentences of context). You might have to toy around with such concepts for your own data. A coffee machine manual it might be OK to only send few sentences at a time, for a novel a longer context might make sense.
Also you might have to experiment with the prompts and the LLMs. Different LLMs might ‘like’ other prompts, or you might try an approach where you let an LLM write the prompt and then send that to the model.