ECO-AIM-RAG-002¶
Name: Excessive retrieval fan-out
Category: AI/ML
Family: Retrieval-Augmented Generation
Primary layer: ai
System layers: ai
Description¶
RAG retrieval queries too many sources, chunks, or indexes before ranking, increasing latency and inference context.
Impact¶
- type: ai-compute
- confidence: 0.65
- notes: Added as part of the 0.3.0 expansion to capture cross-system sustainability and operational waste.
Detection¶
- method: static-or-runtime
- confidence: 0.55
- runtime_validation_required: Yes
Remediation¶
- guidance: Tune top-k, prefiltering, reranking, chunking, and index design against answer quality and cost.
- tradeoffs: May require architecture, product, or operations review rather than a local code change.
Cost Dimensions¶
- compute: high
- memory: medium
- network: high
- storage: medium
- human_time: medium
- carbon: high
- water: medium
Amplification¶
- scales_with_users: Yes
- scales_with_data_volume: Yes
- scales_non_linearly: Yes
Temporal Behavior¶
- startup_only: No
- steady_state: Yes
- burst_sensitive: Yes
- time_degradation: No
Runtime Evidence¶
- retrieval traces
- vector DB metrics
- prompt token counts
Pattern examples¶
No pattern examples provided.
Remediation examples¶
No remediation examples provided.
Metadata¶
- catalog_version: 0.4.0
- status: draft
- source: catalog expansion recommendations applied 2026-05-21