metarank/metarank

What it solves

Metarank is an open-source ranking service designed to make search and recommendation systems smarter and more personalized. It solves the problem of static ranking by allowing developers to integrate real-time user signals (like clicks and purchases) and semantic understanding to optimize for metrics like click-through rate (CTR).

How it works

Metarank acts as a reranking layer that sits on top of existing search engines (such as Elasticsearch or OpenSearch). It processes visitor signals from various streaming data sources and uses machine learning models—including LambdaMART for personalization and MF ALS for recommendations—to rerank result sets in real-time. It also supports the use of LLMs in bi-encoder and cross-encoder modes to improve the semantic understanding of search queries.

Who it’s for

It is built for developers and engineers who want to implement Learning-to-Rank (LTR), semantic search, or personalized recommendation widgets (e.g., "you may also like") without writing extensive custom code for common ranking signals.

Highlights

• Real-time Personalization: Adapts search results based on current user actions and visitor profiles.
• High Performance: Optimized for low latency, handling large result sets within 10-20ms.
• AutoML Capabilities: Features automatic feature generation and model re-training.
• Cloud-Native: Stateless architecture using Redis for state management, allowing for horizontal scaling.
• Built-in Signals: Computes common ranking factors like CTR, User-Agent, and referer out of the box.