Pinecone revealed it is providing early access to a data retrieval engine that understands the underlying intent of a declarative request in a way that enables artificial intelligence (AI) agents to access more relevant data faster.
The Pinecone Nexus engine has two components, with the first being a context compiler that dynamically builds and organizes data on demand for each AI agent based on how an organization operates. A composable retriever then formats and serves responses based on the specific data the AI agent needs to complete a task.
At the core of the capability is a declarative KnowQL query language, which gives AI agents access to six core primitives — intent, filter, provenance, output shape, confidence, and budget — that can be invoked via a single declarative interface.
When added to the managed vector database service that Pinecone provides, it becomes possible to improve accuracy by streamlining the amount of data that needs to be retrieved in a way that also serves to reduce costs, says Jeff Zhu, vice president of product for Pinecone.
The overall goal is to reduce the level of reasoning dependency that AI agents have on legacy retrieval-augmented generation (RAG) workflows that are tied to a specific large language model (LLM), says Zhu. Existing AI agents are stuck in brute-force vector loops. They retrieve a set of chunks, read them, realize something is missing and then retrieve more data until they attain confidence in a response.
Roughly 85% of an agent’s effort is spent on knowledge retrieval, which generally still requires human review before anyone can act on it. “AI agents today are slow, expensive and not especially accurate,” says Zhu.
Instead, Pinecone Nexus shifts that workflow from inference time, when IT teams previously had to consume tokens to run retrieval processes, to the point when retrieval requests are compiled up front, says Zhu. In effect, reasoning is now being moved upstream to improve accuracy and reduce overhead because the AI agent is presented data in a structured format versus just another raw document that lacks any context, he adds.
The composable retriever then serves a curated artifact at query time, including citations and confidence scores, in a way Pinecone claims ultimately reduces the total amount of tokens that might otherwise be consumed by as much as 90%.
Additionally, Pinecone is making available more than 70 applications spanning a wide range of use cases, including reusable instances of requests for product and pricing information to data needed to process insurance claims.
It’s not clear to what degree organizations might ultimately shift more responsibility for reasoning to outside the LLM. One of the primary benefits of that approach is that it makes it less likely that organizations will find themselves locked into a specific LLM provider, notes Zhu.
The one thing that is certain is that data engineering teams should be able to spend more time on data quality issues versus having to create a data pipeline for every use case that requires some type of data to be manually retrieved.
