[Question] React.js serve sentence bert in browser keep reporting models not found.

[bianyuanop]

my codes:

export const useInitTransformers = () => {
  const init = async () => {
    // @ts-ignore
    env.allowLocalModels = false;
    extractor = await pipeline(
      "feature-extraction",
      "Xenova/all-mpnet-base-v2",
    );
  };
  return { init };
};

I'm building a frontend with React that can serve sentence bert directly in browser, but no idea why even i add the line
env.allowLocalModels = false
before pipeline loading the model. In the production environment, it's still trying to access model locally /models/..., but which will never exists in this usecase.

Is there any way i can bypass this check and directly pull the model from remote?

[xenova]

env.allowLocalModels = false should handle this. Could you provide more information about your environment? (OS, browser, build tools, etc.)

[bianyuanop]

Actually we successfully make it work by removing this line, i.e using the default one,
https://github.com/Nexai-app/nexi-v1/blob/main/src/nexai_assets/functions/ml.ts#L52

Still no clue why specify a model doesn't work.

OS: ubuntu 22.04
browswer: chrome 117.0.5938.92
build tools: webpack + rollup

[luweigen]

If you didn't initially set env.allowLocalModels = false, the data read from localhost/models/ will be saved to the browser cache, continuously leading to JSON parsing errors, even if you subsequently set env.allowLocalModels = false. You must manually clear the cache.