/Patent Tracker
NvidiaGranted patentAI & softwareInterest 6/10

Nvidia patents AI memory system that stores documents in neural networks

This filing describes a method for storing documents directly inside a neural network structure rather than in a traditional database. When a question is asked, the system uses the network to retrieve the relevant information and feeds it to a language model to generate an answer.

Why it mattersIt offers a more integrated way for AI systems to access and recall specific facts, potentially making large language models more accurate by tightly linking storage and retrieval.

In the filing’s own words

In various examples, systems and techniques are directed to network-based heteroassociative retrieval-augmented generation (HRAG) for efficient augmentation of inputs into artificial intelligence models. Example techniques include storing documents in a network-based store (NBS) having multiple stages of matrix multiplication(s) and non-linear activation(s). Storing documents includes modifying parameter(s) of matrix multiplications of at least one of the stages. The example techniques further include processing, using the NBS, a query to obtain retrieved document(s) associated with the query and at least approximately reproducing stored document(s). The example techniques further include processing, using a language model, a prompt that is based at least on the query and the retrieved document(s).

What exactly they claim (claim 1)

1. A method comprising: storing a plurality of documents in a network-based store (NBS), wherein the NBS comprises a plurality of stages, an individual stage of the plurality of stages comprising one or more matrix multiplications and one or more non-linear activations, wherein storing one or more documents of the plurality of documents comprises: modifying one or more parameters of matrix multiplications of at least one stage of the plurality of stages; processing, using the NBS, a query to obtain one or more retrieved documents associated with the query, wherein an individual retrieved document of the one or more retrieved documents at least approximately reproduces a document of the plurality of documents; and obtaining a language model prompt based at least on the query and the one or more retrieved documents.

Facts

Company
Nvidia (NVIDIA Corporation)
Official title
Heteroassociative retrieval-augmented stores for AI processing
Patent number
US12743425B2
Application number
18890191
Filed
Sep 19, 2024
Granted
Sep 22, 2026
Inventors
1
Claims
20

Classified under

G06F 16/2455 Physics & computing
Read the full filing (official PDF) ↗ Open in the Patent Tracker

More from Nvidia in AI & software

Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.