AI rules¶
Total rules: 15
Rules¶
ECO-AIM-AI-001 — Oversized model selection¶
Using larger models than needed increases inference cost and emissions.
- Category: AIM
- Family: AI
ECO-AIM-AI-002 — No inference batching¶
No batching increases per-request overhead and lowers throughput.
- Category: AIM
- Family: AI
ECO-AIM-AI-003 — Re-embedding unchanged data¶
Recomputing embeddings for unchanged inputs wastes compute.
- Category: AIM
- Family: AI
ECO-AIM-AI-004 — No prompt caching¶
Repeated prompts without caching waste tokens and compute.
- Category: AIM
- Family: AI
ECO-AIM-AI-005 — Always-on inference endpoints¶
Always-on endpoints waste baseline compute when idle.
- Category: AIM
- Family: AI
ECO-AIM-AI-006 — Unbounded context window usage¶
Excessive context increases token cost and latency.
- Category: AIM
- Family: AI
ECO-AIM-AI-007 — No model quantization¶
Failure to quantize when appropriate wastes inference compute.
- Category: AIM
- Family: AI
ECO-AIM-AI-008 — Re-training without drift detection¶
Training without drift checks wastes compute and introduces risk.
- Category: AIM
- Family: AI
ECO-AIM-AI-009 — No evaluation before scaling model¶
Scaling without evaluation wastes resources and can degrade outcomes.
- Category: AIM
- Family: AI
ECO-AIM-AI-010 — Overly frequent fine-tuning cycles¶
Frequent tuning without clear value wastes compute.
- Category: AIM
- Family: AI
ECO-AIM-AI-011 — Storing all embeddings indefinitely¶
Embedding stores without retention grow unbounded and expensive.
- Category: AIM
- Family: AI
ECO-AIM-AI-012 — Large model in low-SLA workload¶
Using high-cost models where latency/quality needs are modest wastes resources.
- Category: AIM
- Family: AI
ECO-AIM-AI-013 — No GPU utilization monitoring¶
Without GPU utilization metrics, accelerator waste stays invisible.
- Category: AIM
- Family: AI
ECO-AIM-AI-014 — Inefficient feature preprocessing pipelines¶
Preprocessing waste increases training and inference cost.
- Category: AIM
- Family: AI
ECO-AIM-AI-015 — No batching of vector search queries¶
Unbatched vector queries increase overhead and reduce throughput.
- Category: AIM
- Family: AI