vamseeachanta Hybrid Classification Strategy (WRK-1188 Learning)
Hybrid Classification Strategy (WRK-1188 Learning)
For large homogeneous collections, prefer deterministic classifiers over LLM:
| Collection | Strategy | Cost | Accuracy |
|---|---|---|---|
| ASTM (25,537 docs) | Designation prefix → discipline | $0 | 86% vs LLM |
| API/DNV/ISO (1,062) | LLM (Claude Haiku CLI) | ~$2 | Baseline |
| Unknown org (484) | LLM (Claude Haiku CLI) | ~$1 | Baseline |
Rule: If org has predictable title/designation patterns, write a deterministic classifier first. Validate with 100-doc LLM sample. Accept if >85%.
See data/document-index-pipeline skill for full pipeline orchestration.