会社のニュース Context grapher Jedify cuts AI token costs 75 percent

Jedify advocates augmenting AI with context‑graph technology, claiming its solution can slash AI token costs by up to three‑quarters.

Context graphs store entity relationships (people, products, locations etc.) within database‑style structures. Graph databases organise data as nodes (entities) and edges (relationships), prioritising connections for fast traversal of complex interconnected data, unlike table‑based relational databases. While knowledge graphs serve AI workloads, context graphs additionally validate which relationships hold true.

Jedify co‑founder and CTO Adi Elimelech commented: “Much of the industry conflates token efficiency and cost efficiency, yet they differ. Pre‑encoding business logic inside a context graph spares the model from reasoning from scratch for every query; this is the core mechanism, independent of vendor implementation.”

Instead of feeding the full database schema to AI on each query, Jedify’s context‑graph pre‑encodes business logic‑definitions, calculations, relationships and rules‑into a structured knowledge base. It updates and expands autonomously alongside business changes, ingesting new data, relationships and rules. Upon user queries, only relevant entities are retrieved, delivering compact, precise context to the LLM.

Jedify ran production‑warehouse benchmarking across 100 business questions spanning three complexity tiers (200 graded test points total). Its context‑graph setup averaged 25,036 raw tokens per SQL generation call with 87% correct‑answer rate. By contrast, conventional schema‑injection methods consume 50,000‑150,000 tokens per call at 60‑70% accuracy; schema‑injection‑based multi‑agent systems such as CHESS reach nearly 339,965 tokens per request. Note: these reference baselines draw from external studies with distinct schemas and question sets, not side‑by‑side warehouse testing.

Measured by a token‑ROI framework balancing cost and accuracy, Jedify’s approach delivered 4.6× higher returns than schema‑injection workflows and 20× versus basic multi‑agent pipelines. Schema‑injection achieved 60‑70% correct outputs, while Jedify attained strong accuracy for entities within graph coverage.

At the 100‑table enterprise scale, the context‑graph reduced token injection by roughly 50% per SQL call versus raw‑schema baselines; at 200 tables, token savings exceeded 75%.

Elimelech also highlighted data‑ownership benefits: “Business logic resides within your controlled graph, so you do not expose your full schema and join‑filter rules to frontier‑model providers on every request.”

By filtering to only required entities prior to SQL generation, context graphs lower model‑reasoning load. Production‑routing analysis indicates around 85% of enterprise‑analytics queries can run on cheaper open‑source models instead of premium frontier LLMs with negligible accuracy loss, cutting infrastructure spend.

Download the related white paper here.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!