| from __future__ import annotations |
| import argparse,json |
| from pathlib import Path |
| from geomretrieval import GeometricIndex,RAGTop10Config,RAGTop10Ranker,load_beir_zip,load_beir_directory |
|
|
| def load_dataset(path,split): |
| return load_beir_zip(path,split) if str(path).lower().endswith('.zip') else load_beir_directory(path,split) |
|
|
| def main(): |
| p=argparse.ArgumentParser(description='Top-10 RAG shortlist sweep. Deep-recall metrics are intentionally not used for model selection.') |
| p.add_argument('dataset'); p.add_argument('index'); p.add_argument('--split',default='test') |
| p.add_argument('--pools',type=int,nargs='+',default=[25,50,100,200,500]) |
| p.add_argument('--output',default='pool_sweep.json') |
| a=p.parse_args(); ds=load_dataset(a.dataset,a.split); idx=GeometricIndex.load(a.index) |
| out={'dataset':ds.name,'split':a.split,'protocol':'top-10 RAG only','pools':{}} |
| for P in a.pools: |
| ranker=RAGTop10Ranker(idx,RAGTop10Config(pool_size=P)) |
| metrics,_=ranker.evaluate(ds,k=10) |
| out['pools'][str(P)]=metrics |
| print('P=',P,json.dumps(metrics,sort_keys=True)) |
| Path(a.output).write_text(json.dumps(out,indent=2)) |
| if __name__=='__main__':main() |
|
|
