Human setup cost for structured retrieval systems

Determine whether human setup costs—including schema, ontology, and prompt engineering—are material at deployment scale for structured or graph-augmented retrieval systems.

Background

The paper’s cost model quantifies machine costs for embedding, answering, and structured or graph-based indexing, but it deliberately excludes human labor. Establishing whether schema design, ontology construction, and prompt engineering materially affect deployment-scale economics would complete the cost accounting for structured and graph-augmented retrieval systems. The paper explicitly states that its audit does not assess these costs and leaves their materiality unresolved.

References

Our cost model prices machine time (LLM inference and embedding) and not human time. Setting up a structured or graph-augmented retrieval system plausibly involves schema, ontology and prompt engineering by people, none of which we attempted to price and none of which our disclosure audit looked for: that audit searched each paper for currency markers tied to indexing, so it can report the absence of dollar figures but not the absence of any particular cost category. Whether human setup cost is material at deployment scale is an open question this paper does not answer.

— The Commercial Tax: Rent-vs-Own Blind Spots in Multi-Hop Retrieval Benchmarks  (2608.16096 - Sanchez et al., 17 Aug 2026) in Section 5.4, “What a Buyer Should Ask”; Section 6, Limitations