nboost

Why we included this project

NBoost is a proxy that sits between your search client and a backend like Elasticsearch, so you can add neural reranking to an existing stack instead of replacing it. For each query it fetches a larger candidate set than the keyword search would return, then uses a fine-tuned transformer model to reorder those results by relevance. The repo includes ready-to-run models fine-tuned on MS MARCO, including a small TinyBERT variant that runs in roughly 50ms per query on GPU, and deployment is handled through a Docker image or a Helm chart for Kubernetes. If your users keep missing the right documents because plain keyword matching ranks them poorly, this is a fairly direct way to get transformer-based relevance without rebuilding your search backend.