✂️ Open Provence Context Pruner & Reranker

This demo showcases the OpenProvence project, which applies the Provence approach to remove question-irrelevant context while simultaneously computing reranker scores. Modern AI agents and context-engineering pipelines often recurse through large batches of search results. Irrelevant passages inflate the tokens we send to an LLM. By inserting a compact pre-processing model such as Provence, we pass along only the evidence that matters before the LLM generates an answer.

OpenProvence is trained on public datasets with fully open training and inference code, and the weights are released under the MIT License.

Demo Highlights 🚀

• ⚡️ 30M-parameter compact model handles both English and Japanese on CPU -- lightning-fast on GPU.
• ⚙️ Additional model families are available. Swap the model ID to try a different checkpoint.