AI NewsModels & agentsAnnouncement
Nvidia released Agent Knowledge Foundation for enterprise agent retrieval
Nvidia released Agent Knowledge Foundation at GTC Berlin, bundling the Nemotron Retriever open model, an Auto-Ontology library, the Kumo structured-data models and CUDA-X data libraries so an enterprise agent can read documents, map tables to business concepts and make predictions without a separate ETL step.
Image: The New Stack
Why it mattersAn agent that fails on company data has usually failed at retrieval, so a free bundle covering embeddings, ontology mapping and structured-data prediction removes a layer of custom plumbing a team would otherwise write before the agent is useful.
An enterprise agent that is handed a question about a company's own data usually fails at the step before the model runs: find the right documents, map a table to the question, and feed the answer back as grounded context. Nvidia released Agent Knowledge Foundation at GTC Berlin, an open-source bundle built to do that step without a bespoke pipeline per customer.
The collection sits inside the Nvidia Agent Toolkit. Erik Pounds at Nvidia says it includes the Nemotron Retriever open model for embedding, reranking and parsing documents, a new Auto-Ontology library that maps raw tables to business concepts like customers, products and orders, and the Kumo family of structured-data models (Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast). Underneath them run the CUDA-X data libraries cuDF, cuGraph, cuML, cuOpt and cuVS.
The numbers Nvidia published
Nvidia says Nemotron Retriever's embedding model raised search relevance by 65 percent in tests on an open-source OpenSearch setup, with fewer search steps and almost no added latency. Nvidia also says SAP fine-tuned the reranker and ran it nearly three times faster than before, and ServiceNow got better reranking quality while cutting latency by roughly a quarter. Each of these is a figure from a vendor about its own product, and the two partner results are the partners' own tests. The common pattern is a retriever that returns a shorter candidate list, and a reranker that is cheaper to run on it.
Who is building on it
Adel El Hallak, Nvidia's vice president of product for agentic AI, told The New Stack that an agent connected only to one vendor's data store still misses the full picture of a company. The published partner list is long: Palantir, Databricks, Snowflake, Amazon OpenSearch, SAP, ServiceNow and Zoom are using the models in named products, and AWS, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data are building on the foundation.
Zoom runs the Nemotron Retriever models inside Zoom AI Services. Nvidia says the ontology step is what turns a mix of tables into a knowledge graph an agent can read, and the Kumo models then make predictions on top of that graph without a separate feature store or a hand-written SQL pipeline.
A caveat Nvidia's post states on its face. The data stays with the data vendors, not with Nvidia: Snowflake, Databricks and Palantir still own what each agent is allowed to see, which is what makes the single foundation safe to adopt across three catalogues at once. For a team starting a new agent today, this changes which pieces are custom code. Embedding, reranking, document parsing, ontology mapping and tabular prediction are now a bundle that runs on the GPUs the team already has, and the custom part is the business logic the agent actually calls.
Source
NVIDIA Agent Toolkit Expands, Grounding Enterprise Agents in Company Data, Nvidia blog, by Erik Pounds.
Reporting: Nvidia bets open models can give enterprise AI agents the context they're missing, The New Stack, by Frederic Lardinois.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
