Introducing CAFE(S): A framework for defining AI context quality
Atlassian-affiliated researchers introduced CAFE(S), a framework defining five properties of high-quality AI context—clear, actionable, faithful, efficient, and secure—to address rising costs, liability, and security risks in AI-driven systems.
CAFE(S) is a newly published framework in ACM Queue, co-authored by researchers including Atlassian’s Eirini Kalliamvakou and others from Capital One, University of Victoria, and Google, designed to evaluate the quality of context provided to AI systems. Unlike traditional knowledge management, which focused on documentation and peer clarification, CAFE(S) addresses the immediate consequences of poor context in AI workflows, where errors scale rapidly and incur token-based costs.
The framework identifies five core properties of good context: clarity, actionability, fidelity, efficiency, and security. Each property targets specific failure modes—such as stale information undermining fidelity or inefficient context inflating token costs—while security is treated as a distinct concern due to its role in preventing unsafe agent behavior.
Poor context quality has tangible impacts, from financial costs tied to token usage to legal liability, as seen in the Air Canada case where a chatbot’s incorrect refund promise led to a tribunal ruling against the company. Security risks, such as the EchoLeak prompt injection attack, further highlight the need for context that is not only effective but also safe to use in production environments.
CAFE(S) emphasizes treating context as a first-class engineering artifact, requiring ownership, maintenance, and review to ensure it remains accessible, retrievable, and aligned with evolving systems. The framework provides a shared vocabulary for teams to assess and improve context quality, laying the groundwork for reliable AI autonomy without claiming to guarantee success on its own.