Fetching metadata from the HF Docker repository...
-
fix: restore optimized GraphRAG pipeline + 1-10 judge + roberta-large BERTScore
- Restored graph_rag.py: shared gemini_client, adaptive prompting with parametric fallback, entity/doc caps for token efficiency
- Restored judge.py: 1-10 scale scoring, roberta-large BERTScore with rescaling
- Restored llm_only.py + basic_rag.py: accurate Gemini token counts via shared client
- Fixed benchmark.py: HTML report encoding, BERTScore raw fields, datetime fix
- Rebuilt FAISS index from 50+ arxiv papers (185 chunks)
- Results: 8.35/10 judge score, 100% pass rate, 55% graph wins -
Add --num_queries argument to benchmark script
-
Code till Round 1
