Sliding window pipeline extends frozen NER models to long documents

An inference-only pipeline enables frozen transformer models like MahaNER-BERT to perform document-level Named Entity Recognition via overlapping sliding windows. The approach uses Normal or Random Repeat strategies to maintain contextual continuity without requiring architectural changes or retraining.