Query-time retrieval-LLM cost of audited systems

Reconstruct the query-time retrieval-LLM costs of HippoRAG-2, PropRAG, SAG, and KET-RAG, including the costs of entity linking, beam search, and SQL generation, to complete the end-to-end cost accounting of these systems.

Background

The paper measures the answering cost of a fixed gpt-4o-mini configuration over retrieved passages, but not the additional LLM computation performed within each audited retrieval system. Those systems may incur costs for entity linking, beam search, SQL generation, or related query-time operations. Because the source papers do not disclose these figures and the paper’s authors state that they could not responsibly reconstruct them, the end-to-end query-time cost remains unresolved.

References

We measure our own fixed LLM's answering cost, not each audited system's query-time retrieval-LLM cost (entity linking, beam search, SQL generation). None of the four systems disclose that figure either, and we could not reconstruct it responsibly. This is a real gap in the cost accounting, and we state it.

— The Commercial Tax: Rent-vs-Own Blind Spots in Multi-Hop Retrieval Benchmarks  (2608.16096 - Sanchez et al., 17 Aug 2026) in Section 6, Limitations