-
Benchmark: add LLM-as-a-Judge + BERTScore (hackathon 30% accuracy criterion)
Backend (route.ts):
- cosineSim() + rescaleBertscore() helpers for sentence embedding BERTScore
- judgeAnswer() calls LLM with strict PASS/FAIL prompt per answer
- Phase 1 now fetches embed(gold) alongside LLM-only call
- Phase 3 (new): judge(graphrag) + judge(baseline) + embed(graphrag_answer) in parallel
- Aggregate: graphragJudgePassRate, baselineJudgePassRate, avgBertscoreRaw,
avgBertscoreRescaled, bonusJudge (>=90%), bonusBertscore (rescaled>=0.55 OR raw>=0.88)
UI (BenchmarkContent.tsx):
- AggregateData extended with accuracy fields
- New 'Answer Accuracy Evaluation' card with progress bars and bonus indicators
- LLM-as-a-Judge: pass rate %, progress bar, 90% threshold marker
- BERTScore: raw + rescaled, 0.88 threshold marker
- Bonus badge: partial/max/none based on thresholds
- Key Finding card updated to cite judge pass rate + BERTScore
-
Add .gitignore, dataset metadata, retrieval layer, and latest web/graphrag updates
-
🎨 Ultra premium UI: 6 separate pages with animations, bento grid, premium design