NEW YORK, Aug. 26, 2026 (GLOBE NEWSWIRE) — Jedify, the autonomous context graph for data-intensive agentic applications and workflows, today announced the findings of a new benchmark study on context graph architectures, which pre-encode business logic instead of feeding raw database schema to large language models. Among the findings: context graphs can cut AI token costs by up to 75% at enterprise scale while improving SQL generation accuracy.
Jedify conducted the study using its context graph implementation.
The research tested the approach against a live production data warehouse, measuring 100 business questions across three complexity tiers, each run twice for a total of 200 graded data points. The context graph architecture averaged 25,036 raw tokens per SQL generation call and answered 87% of graded runs correctly. For comparison, published research on traditional approaches that inject raw database schema into an LLM reports 50,000 to 150,000 tokens per call and 60% to 70% accuracy, while multi-agent systems built on schema injection, such as CHESS, report figures near 339,965 tokens per request. Those baseline figures come from other studies run on different schemas and question sets, not from a same-warehouse comparison.
Using a Token ROI framework, which weighs both cost and answer accuracy, the researchers found the context graph approach delivered roughly 4X higher return than schema-injection approaches and roughly 18X higher than naive multi-agent pipelines.
“Most of the industry treats token efficiency and cost efficiency as the same problem, but they’re not,” said Adi Elimelech, co-founder and CTO of Jedify, who built and ran the benchmark. “When business logic is pre-encoded into a context graph, the model isn’t reasoning from scratch on every query. That’s the mechanism that matters, regardless of which vendor’s implementation you’re looking at.”
Optimized Token Volume is Just One Advantage
The study illustrates that raw token count alone is an incomplete measure of AI efficiency. Using a token ROI framework, which weighs both cost and answer accuracy, the research found that the context graph approach delivered a 4.6X higher return than schema-injection approaches and 20X higher than naive multi-agent pipelines. Schema-injection approaches returned correct answers roughly 60% to 70% of the time in the study, compared with near-perfect accuracy for the context graph approach on entities it covers.
This efficiency gain also opens the door to lower-cost infrastructure. Because a context graph can narrow each query down to only the entities it needs before SQL generation, the reasoning burden on the underlying model shrinks. The study’s production routing analysis suggests that roughly 85% of enterprise analytics queries can run on lower-cost, open-source models rather than high-end frontier models, without a meaningful drop in accuracy.
“There’s a data ownership piece to this, too,” Elimelech said. “When your business logic lives in a graph you control, you’re not handing a frontier model provider your entire schema and every join and filter rule on every single call.”
The Gap Widens at Enterprise Scale
The study points to independent research to explain why the gap widens as schemas get bigger. Real enterprise data warehouses average 101.5 tables and 869.4 columns, according to the BEAVER benchmark. At that scale, models that reason directly over raw schema tend to lose accuracy: GPT-4o scores 86.6% on the six-table Spider 1.0 benchmark but drops to 10.1% on Spider 2.0, which is built from real enterprise warehouses. Gartner reported in March 2026 that agentic AI workloads consume five to 30 times more tokens per task than standard chatbots, a pattern driving unplanned cost overruns across the industry.
At 100-table enterprise scale, the context graph architecture tested in the study injected roughly 50% fewer tokens per SQL call than a raw schema baseline. At 200 tables, that gap exceeded 75%.
See how context graph architectures can reduce token consumption by up to 75% while improving SQL accuracy—and why the efficiency gap widens as enterprise data environments grow. Download the full benchmark study and read the blog post for the methodology, production results, token ROI analysis and detailed findings behind Jedify’s research.
About Jedify
Jedify is the autonomous context graph for enterprise AI, powering modern agentic applications and workflows with connected enterprise context. By connecting an organization’s structured operational data, including data warehouses, CRMs and financial systems, with its unstructured knowledge, such as documents, playbooks, Slack and meeting recordings, Jedify automatically builds a proprietary context graph powered by Semantic Fusion™. This context graph enables production-grade AI agents and applications that are more accurate, more consistent, and free of the hallucinations and token waste associated with context-free approaches. Jedify is model-agnostic, ensuring enterprises can build on the AI platforms they choose without single-vendor lock-in. For more information, visit www.jedify.com and follow us on LinkedIn.
Contacts
BOCA Marketing for Jedify
jedify@bocamarketing.com
