Conversational Recommender System Using a Combination of Fine-Tuned GPT-4o and Retrieval-Augmented Generation for Laptop Recommendations

Conversational recommender systems (CRS) have revolutionized personalized recommendations in recommender systems by using interactive and adaptive decision-making, particularly in complex domains (e.g., laptops). Existing CRS provides interaction between the system and the user through Form-based Layouts and Natural Language. Natural language-based interactions are typically constructed using Conventional Natural Language Processing (C-NLP) methods. While both interactions have shown certain successes, they also have limitations. Form-based layouts restrict users from expressing their preferences freely because of their rigid and structured nature. On the other hand, C-NLP allows for more dynamic interactions but relies heavily on domain-specific datasets and still struggles to interpret complex user requirements. To tackle these issues, we propose the development of a CRS using Large Language Models (LLMs). Specifically, we combined a Fine-Tuned GPT -4o model and the retrieval technique of Retrieval-Augmented Generation (RAG). LLMs, with their extensive pretraining, can process sophisticated conversations, infer subtle user preferences, and deliver precise recommendations without requiring intensive additional training. At the same time, RAG's retrieval technique effectively incorporates large-scale datasets, ensuring the system maintains relevance and scalability. In evaluation, we compare three models: 1) RAG, 2) Fine-Tuned GPT-4o, and 3) Combined Model (Fine-Tuned GPT-4o + RAG's Retrieval Technique). Results show that the Combined Model excels with an average Hit Rate of 1, Precision of 0.9031, NDCG of 0.9865, and the highest user satisfaction. These outcomes confirm that the approach resolves limitations and ensures scalability.