Abstract:
The current generation of personalised food planning systems does not incorporate individual
cultural eating customs nor community accessible ingredients nor fitness needs-specific dietary
requirements. The present recommendation models depend on global datasets to suggest meals but
they produce unpractical and expensive food choices which neglect Sri Lankan cuisine. The
functionality of modern systems fails to adapt recipes according to fitness goals including weight
loss and muscle gain and dietary restrictions which restricts their effectiveness for fitness
enthusiasts and general users.
To address these limitations, we have developed a Personalized and Culturally Adaptive Meal
Recommendation System using a Retrieval-Augmented Generation (RAG) and Large Language
Model (LLM). The approach is augmented with an in-house recipe embedding model that is
trained on global as well as Sri Lanka-centric recipe databases, capable of supporting meal
recommendations from ingredients. The system pre-processes user-input ingredients using Natural
Language Processing (NLP), converts recipes into vector representations using BERT-based
embeddings, and ranks them using cosine similarity. Further recipe modification layers are used
to ensure that recommendations are aligned with user-specified fitness goals. The model is
deployed using Flask for API management and a React.js-based user interface, offering an
interactive and scalable user experience.
Our system reaches 87% accuracy when recommending contextually and nutritionally suitable
recipes which proves superior to baseline models according to experimental results. The system
demonstrates both precision and recall performance at 78% and 82% which demonstrates its
capacity to create appropriate meal recommendations. The model displays an 8.66 Root Mean
Square Error (RMSE) that demonstrates consistent performance in creating nutritional links to
fitness goals. The research proves that AI-based personalised recommendations will succeed in
providing culturally diverse meals to any user type regardless of their fitness needs.
