load_model_rag.rs

const EMB_MODEL_ID: &str = "sentence-transformers/all-MiniLM-L6-v2";
const EMB_MODEL_REV: &str = "refs/pr/21";

thread_local! {
    static BERT_MODEL: Rc<BertModelWrapper> =   {
        info!("Loading a model on thread: {:?}", std::thread::current().id());
        let model = BertModelWrapper::new(candle::Device::Cpu, EMB_MODEL_ID, EMB_MODEL_REV);
        match model {
            Ok(model) => Rc::new(model),
            Err(e) => {
                panic!("Failed to load the model: {}", e);
            }
        }
    }
}

pub fn get_model_reference() -> anyhow::Result<Rc<BertModelWrapper>> {
    BERT_MODEL.with(|model| Ok(model.clone()))
}