[Digital Today reporter Chi-gyu Hwang] Pinecone, one of the companies that helped pioneer the vector database market supporting retrieval-augmented generation (RAG), is putting out a different message so far on AI agents.

The News Stack recently reported that Pinecone has provided 800,000 developers with chunking, embedding and retrieval methods that support RAG over the past 4 years. It also unveiled its agent knowledge engine Nexus in early May.

Pinecone says retrieval at inference time, or RAG (Retrieval-Augmented Generation), represents the past. It says agents trapped in a retrieve-read-retrieve loop can complete only 50 to 60 percent of tasks, and 85 percent of what agents do is spent fetching context.

Pinecone argues that passing raw chunks to frontier models and hoping the models understand on their own, in other words conventional RAG, is unstable, slow and expensive.

Conventional RAG pulls information and determines meaning at the moment an agent receives a question. Nexus is described as a method that finishes the thinking in advance before a question arrives.

KnowQL, released with Nexus, is a query language for that. Pinecone stressed it can lift task completion rates to above 90 percent and cut token costs by 90 percent through it.

Janakiram MSV, writing for The News Stack, said, "It is hard to take the figures at face value until verified, but the structural argument is headed in the right direction regardless of the numbers. Compile once and read many times is a form that fits agent workloads."

Pinecone is not the only one moving in this direction. Anthropic released frequently used context as reusable bundled skills. Cursor Rules, offered by the AI coding tool Cursor, plays the same role at the editor layer. Claude code sub-agents package context and tools in advance by task. Harrison Chase of LangChain has been calling this 'context engineering' for months. He explained that Pinecone is executing this pattern at the retrieval layer.

There are counterarguments. Janakiram MSV pointed out, "For KnowQL to really show its power, it needs to become an industry-wide standard adopted like SQL. But standards are not created just because one company declares them."