| # | Backend | Discovery Model | Extraction Model | KYS ↓ | Entities | Relations | Graph Size | Knowledge Rate | Total Time | Quality Score | Speed Score |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | gemini | gemini-3.1-flash-lite-preview | gemini-3.1-flash-lite-preview | 62 | 53 | 115 | 6.33 /s | 18.2s | 0.564 | 1.000 | |
| 2 | anthropic | claude-sonnet-4-6 | claude-haiku-4-5-20251001 | 87 | 73 | 160 | 1.75 /s | 91.3s | 0.784 | 0.199 | |
| 3 | openai | gpt-4o | gpt-4o-mini | 47 | 52 | 99 | 1.24 /s | 79.9s | 0.485 | 0.228 | |
| 4 | ollama | gemma4:e2b | gemma4:e2b | 34 | 55 | 89 | 0.67 /s | 132.4s | 0.436 | 0.137 | |
| 5 | ollama | mistral:latest | mistral:latest | 54 | 126 | 180 | 0.67 /s | 268.2s | 0.882 | 0.068 | |
| 6 | ollama | gemma3:12b | gemma3:4b | 62 | 45 | 107 | 0.55 /s | 194.0s | 0.524 | 0.094 | |
| 7 | ollama | gemma4:latest | gemma4:latest | 85 | 72 | 157 | 0.53 /s | 297.4s | 0.770 | 0.061 | |
| 8 | ollama | qwen3.5:9b | qwen3.5:9b | 99 | 105 | 204 | 0.51 /s | 400.7s | 1.000 | 0.045 | |
| 9 | ollama | qwen2.5:14b | qwen2.5:3b | 37 | 30 | 67 | 0.41 /s | 163.3s | 0.328 | 0.111 | |
| 10 | ollama | llama3.2:latest | llama3.2:latest | 29 | 36 | 65 | 0.40 /s | 163.4s | 0.319 | 0.111 | |
| 11 | ollama | qwen3:8b | qwen3:8b | 58 | 66 | 124 | 0.38 /s | 329.4s | 0.608 | 0.055 | |
| 12 | ollama | llama3.1:8b | llama3.1:8b | 38 | 46 | 84 | 0.33 /s | 258.5s | 0.412 | 0.070 | |
| 13 | gemini | gemini-3-flash-preview | gemini-3-flash-preview | 13 | 10 | 23 | 0.30 /s | 76.7s | 0.113 | 0.237 | |
| 14 | ollama | glm-4.7-flash:latest | glm-4.7-flash:latest | 62 | 66 | 128 | 0.27 /s | 466.1s | 0.627 | 0.039 | |
| 15 | gemini | gemini-3.1-pro-preview | gemini-3.1-pro-preview | 0 | 0 | 0 | 0.00 /s | 204.3s | 0.000 | 0.089 |
The Knowledge Yield Score (KYS) is a composite metric that balances output richness (how much the pipeline extracted) against efficiency (how fast it ran). It is the geometric mean of two normalized sub-scores, analogous to the F₁ score — it penalises runs that excel in only one dimension.
quality_norm rewards pipelines that produce large,
dense knowledge graphs. speed_norm rewards pipelines
that finish quickly; the fastest run scores 1.0 and slower runs are
penalised proportionally. The geometric mean ensures a pipeline
cannot compensate for poor speed with high quality alone — both
dimensions must be strong for a high KYS.
Knowledge Rate (graph_size / total_time) is a complementary raw throughput metric expressed in graph
elements per second, useful for absolute comparisons independent of
the normalization range.