Using LLMs to Build Persistent Preference Profiles for Better Recommendations

Summary

A writer explored whether a large language model could learn a person’s detailed tastes instead of relying mainly on search history or broad genres. After spending about two hours describing preferred plot structures, characters, humor, and other literary details to ChatGPT, the writer received five book recommendations that were unfamiliar but appealing. The model also correctly predicted whether the writer would like several previously read books and explained its reasoning, although the experiment had a very small sample. The writer then built a simple prototype that stores a persistent preference profile alongside user conversations and uses conversational LLM functions to extract and update preference information over time. The proposed architecture uses the evolving profile to guide later recommendation searches, rather than treating each recommendation request as an isolated prompt. The author suggests that profiles and conversations could remain on the user’s device or in a cloud chosen by the user, and wonders whether data could be minimized or pseudonymized before being sent to an LLM. The possible privacy benefits of sending less information to search services are presented as an unverified hypothesis. The author also asks whether the approach would work beyond books, whether other people could articulate preferences well enough, whether users would adopt it, and whether the overall architecture is novel.