An agentic LLM-powered recommendation system using tools, templates and prompt engineering

Cold start and personalization remain fundamental challenges in recommender systems, particularly in e-commerce and entertainment platforms where new users and items lack historical interaction. Our paper proposes a cross-domain, Large Language Models (LLMs)-powered recommendation framework that mitigates cold start by shifting from interaction-based learning to semantic understanding and reasoning over heterogeneous textual and contextual information. The system integrates three embedding models (all-MiniLM-L6-v2, all-mpnet-base-v2, multi-qa-MiniLM-L6-cos-v1) for semantic retrieval with two open-source LLMs to generate adaptive, explainable recommendations through natural language interaction. An agentic filtering mechanism dynamically incorporates user constraints such as category, usage scenario, price range and ratings, while prompt templates include user attributes and prior purchases to enhance personalization. The approach combines semantic similarity, reasoning-driven recommendation powered by a local model llama3.1:8b or Groq-hosted llama-3.3-70b-versatile model and content-based signals within a unified architecture. The embedding models are evaluated using ranking and retrieval metrics (Precision@K, Recall@K, NDCG@K, Mean Reciprocal Rank, Hit Rate@K, catalog coverage). Experimental validation is conducted across two distinct domains: retail products and movies, using publicly available datasets from Hugging Face and Kaggle platforms. The proposed framework is dataset-agnostic, explainable by design and readily transferable across domains, offering a scalable solution for cold start-aware recommendation systems.