This project explores how LLM-based conversational agents can be integrated into real-world recommender systems without replacing classical ranking models.
Rather than treating the LLM as a black-box decision maker, the system is designed around a clear separation of responsibilities:
-
Ranking models remain responsible for relevance and optimization
-
The LLM acts as an intent interpreter, translating unstructured user preferences into explicit, validated constraints and ranking boosts
-
A retrieval-augmented grounding layer (RAG) ensures that conversational signals are tied to observable product attributes and review evidence
The project is intentionally scoped to remain realistic, measurable, and defensible, prioritizing design choices that could plausibly work in production environments.