Thank you. Those are good points about the applicability of retrieval-augmented generation and about there being potentially more than one embedding space involved.
Wisdom materials may have multiple interpretations and thus uses. For instance, the proverb “a rolling stone gathers no moss” has two popular interpretations: (1) people pay a price for being always on the move, in that they have no roots in a specific place (the original meaning), and (2) people who keep moving avoid picking up responsibilities and cares.
Without personalization, conversational search and recommender systems could utilize interactive dialogical processes to help end-users explicate their scenarios or queries to best sift and sort the space of wisdom materials.
With personalization, we might envision there being two input vectors. One would be a vector for a model of the particular end-user. The other would be a vector for their provided scenario or query. It should be possible to combine or to otherwise transform these two input vectors into a larger vector in a higher-dimensional embedding space.
Regarding modeling end-users with vectors, using the life2vec
approach [1], or something like it, end-users could be modeled by means of using sequences of life-events.
We can also consider recommender systems for wisdom materials which present end-users with opportunities for providing feedback. After receiving their results, end-users could then click upon, or otherwise upvote, those wisdom materials which were most useful.
References
[1] Savcisens, Germans, Tina Eliassi-Rad, Lars Kai Hansen, Laust Hvas Mortensen, Lau Lilleholt, Anna Rogers, Ingo Zettler, and Sune Lehmann. “Using sequences of life-events to predict human lives.” Nature Computational Science (2023): 1-14. [PDF]