Cost-Governed RAG Enables Unified Per-Tenant Cost Attribution

Enterprise deployments of Retrieval-Augmented Generation (RAG) systems often struggle with accurately attributing costs across different tenants, particularly for the retrieval layer which includes vector memory and embedding API calls. This paper introduces Cost-Governed RAG, an innovative architecture designed to provide precise, unified per-tenant cost attribution for both retrieval and generation components. The core of this architecture is the integration of TurboVec, a codebook-oblivious vector index, with a multi-tenant LLM governance gateway. TurboVec's deterministic memory formula allows for near-exact calculation of per-tenant retrieval costs, a capability not typically found in graph-based indexes. Deployed within a cloud data platform, the system demonstrated 99.96% end-to-end cost attribution accuracy across simulated tenants, with minimal telemetry overhead. Furthermore, it achieved substantial reductions in retrieval infrastructure costs, ranging from 3.1 to 9.0 times compared to managed vector database services, while also enhancing security by removing shared-codebook leakage surfaces.