feat: increase max file upload limit to 100MB

feat: add LibreOffice support in Dockerfile and simplify extractor.py for legacy doc/ppt conversion

Convert tumpang tindih sentence similarity to pure Jaccard Index, removing all TF-IDF and Cosine dependencies from overlap calculation

Update main overlap_score calculation to use pure Jaccard Index of stopword-filtered keyword sets

Implement Jaccard index stopword and conjunction filtering in compute_overlap function

Restore weighted ranking selection in BERT backend while preserving raw cosine similarity scores for frontend UI display

Clean up project structure: delete model_lexical.py and remove all its references/fallbacks, move compute_overlap to preprocessor.py, and remove TEKS COCOK badge from sidebar, retaining only sentence snippet previews

Fix summary score calculation (remove source_weight multiplication) and add real-time search filtering & text-matching in the sidebar & summary panels

fix: upgrade torch to 2.5.1+cpu — fixes torch.distributed.tensor.device_mesh import error with latest transformers

fix: simplify topic name to 'Topik N', add bert_error field to API result for debugging

fix: upgrade torch to 2.4.0+cpu (required by sentence-transformers>=3) and pin numpy<2 for binary compat

perf: use torch CPU-only wheel (~200MB vs 1.5GB) to speed up HF Spaces rebuild

refactor: simplify BERTopic pipeline to exact LOTO config from notebook (no fallback scenarios)

chore: align model print statement with BERT model backbone

fix: implement robust BERTopic fallback search to resolve empty topic results and 2.0 scores

Fix scipy linalg eigh crash on small datasets by implementing dynamic UMAP/HDBSCAN parameter scaling

Update model to snickerdust/FT-PLN-IndoBERT and apply optimal LOTO parameters

add BERTopic support for topic modeling

upload ke HF