For the last three years, the default engineering response to any enterprise AI context problem was automated: “Just build a RAG pipeline.”
… But as we advance through 2026, the cracks in the RAG-everything approach have become too wide to ignore. Chunking destroys complex table structures, vector retrieval is inherently probabilistic (you might get the right chunk, or you might get an outdated one), and keeping embeddings synchronized with rapidly updating data is an absolute operational nightmare.
To solve this, Google Cloud quieted the “RAG-everything” noise by open-sourcing the Open Knowledge Format (OKF v0.1). It isn’t a new cloud database, an LLM framework, or an SDK. Instead, it is a vendor-neutral, portable specification that formalizes the “LLM Wiki” paradigm — the exact structured, interconnected “brain” concept long advocated by AI researchers like Andrej Karpathy. — Read More
Daily Archives: July 15, 2026
The Full AI Stack Explained: Every Layer, Every Tool, and How to Make the Right Choice at Each One
Most people building with AI are one or two layers deep in a seven-layer system.
Every week someone asks me what tools they should use to build their AI system. The answer is always the same: it depends which layer of the stack you are talking about — and most people asking the question have not The Full AI Stack Explained: Every Layer, Every Tool, and How to Make the Right Choice at Each One — Read More
Evals Are the New OKRs
[E]vals are becoming essential for AI.
…Organizations need metrics that show whether work is producing the intended result.
OKRs should not be activity lists.
Key results should not simply measure work completed.
They should define visible, tangible outcomes.
… If an organization cannot define what good AI output looks like, AI will create more output without necessarily creating more value.
Evals, metrics, and key results all serve the same deeper purpose. — Read More
Rick's Cafe AI 