Skip to content

Search-R1

Train language models that reason and call search engines

Visit official site

What Search-R1 is

The framework is described as a reinforcement learning framework built on veRL for training language models that reason and call search engines.

What you can do with it

  • Train an LLM to reason and call a search engine
  • Use local sparse or dense retrievers and online search engines
  • Run multi-node training for LLMs above 30B
  • Define custom datasets and custom search engines

Application scenarios

  • Train a model that answers complex questions with search engine calls
  • Build a question answering system with multi-turn dialogue and live retrieval
  • Apply it in education to widen a learning assistant with search

Model notes

  • Several reinforcement learning methods are supported, including PPO, GRPO and reinforce
  • Multiple LLM architectures are supported, including Llama3 and Qwen2.5
  • The project is stated to be open source
  • The stated focus is efficient and scalable tool-augmented reasoning research

Pros & cons

✓ What we like

  • Local sparse or dense retrievers and online search engines can both be used
  • Custom datasets and custom search engines are supported

! What to watch out for

  • Training relies on reinforcement learning methods configured in the framework
  • The documented setup needs an index, a corpus and a local retrieval server

FAQ

What is it built on?

The material states the framework is built on veRL and supports several reinforcement learning methods and LLM architectures.

What are the documented starting steps?

Install dependencies, download the index and corpus, process the training dataset and start a local retrieval server.

Last reviewed: 2026-11-15

More AI search engine tools

View all →

How we review