AI DATA PIPELINE Aug 31, 2026
Blocks & Files covers (Aug 27) Jedify's claim that pre-encoding business logic into a context graph cuts AI token costs by up to three quarters. Instead of exposing an entire database schema to the model on every query, the graph pulls only the handful of entities relevant to the question. Jedify's own testing measured ~25,036 raw tokens per SQL-generation call at 87% graded accuracy, versus 50,000–150,000 tokens at 60–70% accuracy in published schema-injection baselines, and ~339,965 tokens per request for the CHESS multi-agent approach. Caveat stated by the outlet: those baselines come from other studies on different schemas, not a same-warehouse comparison — treat the 75% figure as vendor-measured, directionally plausible.
The admin lesson: the AI bill is increasingly a data problem, not just a model problem — token spend scales with how much irrelevant context you ship to the model, and retrieval quality (not raw data volume) decides it. That puts storage and data teams back in the critical path: dedupe and govern the estate, keep metadata clean, and design the data pipeline (ingest, transformation, embedding/vector indexing) as deliberately as the model layer. NetApp's AI Data Engine and BlueXP AI/ML positioning target exactly this pipeline, and the same discipline applies to ONTAP estates today: checkpoints, model artifacts, and RAG corpora are just files — snapshot and dedupe them like everything else.