Spaces:
Sleeping
Sleeping
Commit History
-
Add notebook results and fresh embeddings cache
Dipanshu Singh +1: Claude Opus 4.6- Notebook run with full evaluation outputs (16,407 docs) - Retrieval: Precision@3 0.67, Recall@3 0.75, MRR@3 0.69 - Generation: ROUGE-L 0.51, BLEU-4 0.00 - Fresh embeddings cache for 16,407 Q&A pairs Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>e50ea49
-
Rebuild CSV from original MedQuAD XML with source URLs
Dipanshu Singh +1: Claude Opus 4.6- Parse all 12 MedQuAD XML collections from github.com/abachaa/MedQuAD - Extract Document-level url attribute (was missing in previous CSV) - Filter to 16,407 Q&A pairs with non-empty answers - 100% valid source URLs (was 0% before) - Remove stale embeddings cache (will recompute on next run) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>a0c7680
-
Add pre-computed embeddings cache and optimize load_pipeline
ClaudeUnverified signature- Commit embeddings_cache.npy (33 MB, 10979×768 float32) so Streamlit Cloud loads instantly instead of computing embeddings from scratch at every startup - Increase encode batch_size from 16 → 128 for faster CPU/GPU throughput - Validate cache row count against CSV to auto-recompute if data changes - Use model.encode() directly instead of manual batch loop (cleaner, same result) https://claude.ai/code/session_015CkjMVc96oHM4F5TjbbTr26384c6b