Philosophy
The thinking behind CoreRec.
One API, Every Model
Recommendation systems span decades of research: matrix factorization, neural networks, graph methods, sequential models. Most frameworks force you to learn a different interface for each. CoreRec treats every model as a recommender. Same fit(), predict(), recommend(). Swap SVD for DCN without rewriting your pipeline.
Research to Production
The gap between a Jupyter notebook and a deployed system is where projects die. CoreRec is built so the code you prototype with is the code you ship. Retrieval, ranking, reranking compose into pipelines. Models persist and load. A lightweight serving layer handles inference. No separate "production" fork.
Composability Over Monoliths
Real systems are multi-stage. Retrieve candidates, score them, rerank for diversity or business rules. CoreRec exposes each stage as a component. Mix collaborative retrievers with semantic search. Add fairness or diversity rerankers. The pipeline is the product of its parts, not a black box.
Breadth Without Bloat
57+ models across deep learning, matrix factorization, graph, sequential, and Bayesian families. The goal is not to own every paper. It is to give practitioners a single place to experiment, compare, and choose. Type-hinted, documented, tested. Pick what fits your data and latency budget.
Elegant by Default
CoreRec is the application of deep learning to recommendation systems that establishes a unified framework for research and production. The philosophy is simple: reduce friction, preserve flexibility, and let the models do the work.
Creator
Vishesh
CoreRec is built by Vishesh, who works on recommendation systems and ML infrastructure. The project started from the need for a single, consistent API across research and production. Open source, maintained in spare time.
GitHub