The answer correctness score is assigned as an integer value between 1 and 5 by gpt-3.5-turbo. After scoring the entire evaluation dataset, min-max normalization was applied to scale the scores between 0 and 1.
RAG Flow | Answer correctness score |
---|---|
baseline | 0.658 |
baseline + hyde | 0.676 |
baseline + query decompose | 0.397 |
baseline + hyde + bm25 | 0.750 |
baseline + hyde + hybrid | 0.768 |
baseline + hyde + colbert rerank | 0.787 |
baseline + hyde + hybrid + cohere rerank | 0.820 |