AI-Assisted Knowledge Access for Legacy Enterprise Asset Management in Energy Operations: A Practical Retrieval System
This paper presents a practical AI-assisted retrieval system that enhances knowledge access in legacy energy enterprise asset management platforms by combining intent understanding, semantic enrichment, and hybrid retrieval techniques, demonstrating significant improvements in retrieval precision, task completion time, and user confidence during a pilot study.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are walking through a massive, ancient library that holds the secrets to running a power plant. This library isn't organized like a modern one with clear labels and a digital search bar. Instead, it's a chaotic maze of dusty scrolls, handwritten notes, and blueprints from fifty years ago. Some books are written in a strange code only the oldest librarians understand, while others are just lists of numbers with no explanation. In the world of science and engineering, this is what happens when companies rely on "legacy systems"—old software that has been running for decades. It's too expensive and risky to throw it away and build a new library, so they have to keep using the old one. The big question researchers are asking is: How do we help people find the right information in this mess without tearing the whole building down? The answer lies in a concept called "Retrieval-Augmented Generation" (or RAG for short). Think of RAG as a super-smart, hyper-organized librarian assistant. Instead of just guessing the answer or making things up, this assistant is trained to run to the specific shelf, grab the exact page, read it, and then tell you the answer while pointing exactly to where it found the proof. It combines the creativity of a storyteller with the strict accuracy of a fact-checker.
This paper tells the story of a team that built exactly this kind of librarian assistant for Ontario Power Generation, a company that manages energy operations. They faced a common problem: their engineers were wasting hours every day trying to find answers in their old, clunky computer systems. The information was there, but it was buried under confusing names, scattered documents, and a user interface that felt like navigating a spaceship without a map. The team created a "retrieval assistant" designed to act as a practical overlay—a layer of intelligence sitting on top of the old system to make it usable again.
The assistant works by understanding what the user is actually asking, even if they use the wrong words. For example, if an engineer asks, "Where do I fix the pressure valve?" the assistant knows to look for specific technical terms and old codes that the system uses. It then searches three different "families" of information: the original manuals from the machine makers, the hidden database structures (which are like the skeleton of the system), and the company's own specific rules and how-to guides. Once it finds the right pieces of information, it doesn't just dump a wall of text on the user. Instead, it builds a clear, grounded answer that includes direct links. If the answer mentions a specific screen or panel, the assistant turns that mention into a clickable link that takes the user straight there, saving them from hunting through menus.
The results of their pilot test were quite promising. When they compared the new assistant to the old way of doing things (where engineers had to search manually), the difference was clear. The assistant improved the quality of the answers it found. Specifically, when looking at the top five results the system suggested, the accuracy jumped from 0.56 to 0.72. It also got better at putting the right answer at the very top of the list, with a score called "Mean Reciprocal Rank" rising from 0.43 to 0.58. Perhaps most importantly for the engineers' daily lives, the time it took to finish a task dropped significantly. The median time to complete a task fell from 14.2 minutes down to 8.3 minutes. That's a reduction of about 41.5%.
The engineers who tested the system also felt more confident and found the tool more useful, rating it 4.0 out of 5 on both scales, compared to lower scores before the tool was introduced. The system also became much better at avoiding mistakes; the rate of failed or useless answers dropped by half, from 0.14 to 0.07. However, the authors are careful to note that these findings come from a small pilot study with a limited number of participants and prompts. They describe the results as "pilot findings" rather than a final, proven solution for every situation. They suggest that while the method works well in this specific environment, more testing with larger groups is needed to be sure it holds up everywhere.
The paper also highlights what the system doesn't do. It doesn't replace the old software or rewrite the company's history. It doesn't magically fix bad data or outdated manuals; in fact, the authors found that the system works best when the underlying information is carefully prepared and "enriched" with clear descriptions. If the source documents are vague or the terms are confusing, the assistant can still get stuck. The team explicitly ruled out the idea that a fancy AI model alone could solve the problem; instead, they found that the secret sauce was "semantic enrichment"—a process of cleaning up and labeling the data so the AI could understand it better.
In the end, this paper suggests that you don't always need to demolish an old building to make it useful again. By adding a smart, guided layer on top, you can help people navigate the past while they build the future. The team hopes this approach will become a standard way to handle old systems in energy, banking, and other industries, turning confusing archives into accessible knowledge bases. They plan to keep testing the system, tracking how it performs over time, and refining the rules to make sure it stays accurate as the company's data changes. For now, the evidence suggests that a little bit of AI guidance can make a very old system feel brand new.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.