Hitachi promises time travel for LLMs – just don’t lose your RAG

Hitachi Vantara is claiming to have cracked the problem of time travel in AI, though its biggest immediate challenge seems to be whether to charge for it.

The vendor unwrapped its Hitachi iQ Time Machine technology at Computex in Taiwan last month. While the concept appears to be bending the laws of physics, it’s actually more about applying the lessons of version control to LLMs.

In a video detailing the feature, Hitachi Vantara says today’s AI systems “are caught in the present” because current LLMs “rely on documents and data that are routinely updated and replaced. Previous versions gone forever.”

It claims to have delivered “Time Aware AI” through Time Machine “Powered by Blackwell GPUs, deep reasoning NIMS, and NEMO retriever,” which it claims allows users for the first time to “access data from different time periods.”

Hitachi’s agent accesses documents on Hitachi VSP One Object, meaning they are protected by enterprise grade security and privacy controls, while its versioning means older as well as the newest versions are accessible.

Being able to combine old and new versions of documents might not sound particularly groundbreaking. But it does address some very real world problems enterprises face when working with LLMs, Hitachi CTO for AI Jason Hardy said, pointing out that “AI has no sense of compliance.

“AI does not understand how data changes. It just knows about today’s version of it or when it was last looked at. So what we were able to do with that is bring across data [and] also introduce the concept of versions of data, how it changes over time.”

Hardy explained the aim of Time Machine was to “provide RAG capabilities with enterprise compliance in mind.”

He said: “This is a RAG-type feature that includes the VectorDB, as well as linkage into the Hitachi VSP One Object platform, and the LLM necessary to interact with the content.

But unlike traditional RAG implementations, Time Machine “understands how data changes over time and allows customers to roll back the LLM’s point of view to different points in time, aligning with how the data captured has changed over time.”

When it comes to changing or un-embedding data in LLMs, he said, the company has created IP that “allows for time to be associated with embeddings into the VectorDB.” He described these as “temporal embeddings. With this capability, we now can ‘activate’ previous versions of data through the time aware embeddings, as well as blacklist/remove data entirely.”

This is about more than a couple of hundred documents, he said. “You’re doing buckets of data to provide that value.”

More time embed

The user, or the application or API call, can then tell the system, “Wait a minute, the data that I’m seeing today, from a results perspective, something doesn’t match than what I was expecting and what I got a week ago.”

Users can then roll back and run a query, and, ask the model a question on data that’s a week old or a year old. “Through that, we now can say, Okay, this is what your drift looks like. This is what your data looks like. AB, compare it.”

If it became clear that something had gone awry with the data in the interim, that would have previously “required completely re embedding all your data. You would have to delete it and start over. This actually allows you to roll back pieces of the data in the model to previous versions.”

Hardy said this had obvious applications when sensitive customer data is introduced into a model. “I can now undo that individual document being embedded into the LLM, I can undo whatever created that customer data and how it was brought into the model.” Hardy said the technology was very early stage. “We’re still looking at how we market it. Do we give it away for free? What does that actually mean? How do we embed it in things like that.”