Work RAG and LLM systems plate 02 / 16

Recipe Retrieval at 500K Scale

Course project · University of Oulu · Oct - Nov 2025

500K+ recipes · NDCG@k 0.95 · MRR 1.0

A course project at the University of Oulu: large-scale retrieval over a corpus of 500K+ recipes, with a RAG layer on top.

The system combines BM25, Sentence-BERT embeddings and hybrid search, built on an end-to-end NLP pipeline: text preprocessing, semantic indexing, structured nutritional filtering and query expansion.

I evaluated BM25, semantic and hybrid retrieval against each other using standard IR metrics. The system reached NDCG@k 0.95 and MRR 1.0.