BGE-Reasoner: An open-source framework for reasoning-intensive retrieval

Hi HN, we're sharing an update on our work to improve reasoning in retrieval systems for RAG and Agents.

Our project, BGE-Reasoner, is an open-source, three-stage framework (Rewrite, Embed, Rerank) that showed strong performance on the BRIGHT benchmark for reasoning-intensive retrieval as of our submission on Aug 21.

Our main contribution is using synthetic data and reinforcement learning to handle complex queries that go beyond simple semantic matching. We're sharing this because we believe the framework itself is a solid, replicable contribution for anyone working on advanced RAG or Agent search.

We're in the process of open-sourcing the model weights, code, and training data to the community.

Love for you to take a look, share your thoughts, and we welcome any and all feedback or critiques. Thanks!

- Github: https://github.com/FlagOpen/FlagEmbedding/tree/master/resear...

- Benchmark: https://brightbenchmark.github.io