KATIA-v1 SERIE, DEVELOPMENT FROZEN IN v0.9.55

Long story short: both Gemini and Kimi converged on the same RAG-oriented
management of the past interactions (faster and more efficient) but they both
also drop ever session persistence adopting the single-turn mental model (aka
Alzheimer mode)

It remembers the past but not why it went in the kitchen and this disrupts some
specific commands of Katia plus a reasonably good way to debug it further soon
after finally it was working pretty well! (enough to catch fuffa-guru in action
on the AI hype).

=-> https://github.com/robang74/chatbots-for-fun/
    commit/e1f2071f3321179d4aa64beabc34c61dbed7b47e (freeze point)

Since the "WHEN TO RAG? Always!" a suggestion has been implemented also in
Google Gemini, all the previous interactions (chat session information) are
provided to the next prompt in read-only mode as if that prompt were the first
one.

On the other hand, caching information between-prompts is not an accessible
feature anymore. Under this perspective [CSC] is provided by a sort of RAG
about all the previous interactions while the rules on how to use it are still
working. While the concept of IBPS (persistence between prompts) fell apart,
soon after it started to work as expected. Which brought commands like
'show-savings' and 'print-ops' to show 'none'.

Questioning the AI about these failures is almost useless because the
explanations are created as post-hoc artifacts (aka hallucinations). In fact,
these explanations are not consistent in terms of checking. The AI fails
"smoothly" to execute some part of the prompt and it has no idea of the root
cause of those failures.

Hence, development is broken in its essence, especially because without
persistence between prompts, debugging becomes extremely difficult. In fact, in
some terms every prompt is the first prompt but with more RAG. So, debugging
what? A sort of Alzheimer for which short-memory is gone?

CONCLUSION

The "American Dream" is not accessible by Italy.

================================================================================

WHEN TO RAG? BY GUY ERNEST

=-> https://github.com/guyernest/advanced-rag (github)

Always: another problem solved.

Retraining a LLM without the cost of retraining:

=-> https://robang74.github.io/chatbots-for-fun/html/
    gemini-context-retraining-for-human-rights.html (article)

=-> https://arxiv.org/pdf/2507.16003 (paper)