Apply NEF and FFF for RAG

[MrCsabaToth]

Named Entity Filtering (https://blog.cubed.run/eliminating-hallucinations-lesson-1-named-entity-filtering-nef-5f5956d748e0) and Fully Formatted Facts (https://medium.com/@JamesStakelum/the-end-of-ai-hallucinations-a-breakthrough-in-accuracy-for-data-engineers-e67be5cc742a) along Noun-Phrase Dominance supposed to greatly decrease hallucinations and increase RAG performance (in terms of preciseness and correctness).

I've seen even earlier BERT-based and other few models which were able to generate tags for chunks automatically. These tags should be stored as meta-data for the RAG and should be utilized for filtering to avoid problems depicted in the above articles.

We should also tune the resolver prompt we use for conversation piece embedding preparation to include instructions about "avoid any ambiguity" besides resolving / unfolding contextual references.