Samuel is tasked with investigating voltage fluctuations at a small hydroelectric dam. Upon arrival he takes a look at the voltage regulator, then tethers his laptop to his phone in order to access the system of record and understand the asset's history. After 30 minutes of searching, he identifies a pattern of regular fluctuations over the past year, each time addressed by a different electrician. Samuel then goes on a hunt for the voltage regulator's manual on the company shared file drive. That task takes another half hour. Finally, he spends another hour skim-reading through the 300-page manual for the device. Only after nearly two hours of preparation does Samuel begin the actual work of rectifying the issue.

Why now?

Samuel's experience highlights the inefficiency of asset information dispersed across multiple systems. This calls for a better solution. Advancements in AI technology make it feasible to provide such a solution.

• Knowledge Fusion: With digital systems of record and document management systems accessible via API, it's now technically possible to aggregate all relevant information about an asset.
• Transformer-based Machine Learning: The Transformer neural network architecture, underlying LLMs like ChatGPT, can learnt to find the specific relevant information within a large document set. This capability allows for efficient processing of large volumes of data to quickly and accurately answer specific questions.
• Vector Database: While LLMs are powerful, they struggle with digesting extremely large datasets. A well-configured vector database enables the extraction of specific insights from datasets of virtually unlimited size.

How does it work?

This solution involves several steps:

• Integration with Systems of Record: Integration can be either real-time or asynchronous, with the latter being achieved through methods like FTP or web uploads. This flexibility ensures immediate value from the system, with deeper integrations providing benefits of the most up-to-date information.
• Data Collection and Processing: A combination of traditional query and vector search methods is employed to collect the relevant subset of information.
• Utilizing LLMs for Insights: The gathered data is then fed into an LLM with a carefully crafted prompt, providing sufficient context for accurate response generation.

Remaining Challenges

Despite these advancements, several challenges still need to be overcome:

• Limitations in Question Answering: LLMs may struggle with questions requiring multiple steps like e.g. statistical analysis. In such cases, the LLM can output structured data for subsequent external statistical evaluation, but they cannot directly answer the question.
• Risk of Hallucinations: There's a risk of LLMs generating inaccurate information, though efforts are underway to minimize this by enhancing the model's ability to indicate uncertainty.
• Dialog Speed: While the process provides rich insights from multiple data sources, it is often slower than traditional search because of the slow speed of the data fusion process combined with limitations inherent in the LLM’s text processing.

How Datch solves this

The hypothetical scenario showing Samuel’s experience shows the challenge of understanding an assets history and documentation. By integrating advanced technologies like LLMs, knowledge fusion, and vector databases we can make it easy to ask questions about an asset’s history. This can significantly reducing the time and effort required for tasks like Samuel's, and paving the way for more efficient and informed decision making.

Below is a sneak peek of how it appears in the Datch UI.

‍

Want to see it in action for yourself? Let us know by dropping ur email in the box below.

‍