Delta
The AI engine empowering all of our products.
Overview
Delta is the foundation AI engine that enables smart features across all of our products.
As companies integrate centralised data repositories, larger amounts of information are available for analysis. What if we could allow our customers to leverage this knowledge, in a way that's simple and safe? That's the entire goal behind Delta: whenever our products come up with questions and associated context, Delta works out answers in a segregated environment and returns them.
Technically speaking, Delta leverages RAG-powered SQL generation to distill relevant insights. You can further explore its architecture here.
Background
Many companies are starting to collect of their crucial data in one or more centralised data repositories. While this vast amount of information has the potential to unlock insights on the business itself, it is often untapped. Leveraging it requires:
- proficiency in how to query the data
- in-depth understanding of the specific data architecture.
How do decision makers deal with data-driven decisions?
It's usually one of two ways:
- Existing reporting tools or dashboards. Besides the effort it takes to build and maintain such systems, these are often sub-optimal: they force a pre-built template on any question.
- Manual extraction from engineers or data specialists who are able to efficiently query companies data to retrieve the requested information. This is not an efficient solution either, since these experts are often overloaded with work and can't be responsive.
Throw in the mix the inherent looping of this process - once an answer is provided, it often raises additional questions - and you get a frustrating experience on all sides. Decision makers either can't get critical information or wait for a long time, engineers are interrupted from their core tasks.
Delta is our answer to this problem for every knowledge domain related to our products:
- Its natural language interface allows everyone to ask questions, completely removing the need for SQL expertise.
- Users can keep asking questions freely as long as needed, promoting exploratory data analysis.
- More people can access data-driven insights, encouraging a culture of informed decision-making across the entire organisation.
Architecture
Delta is a RAG system that uses multiple LLMs to extract and interpret data.
It includes:
- A Vector DB with tenant segregation, plus a RBAC system that ensures that each submitted input is enhanced only using data from the customer making the request.
- A LLM that has been finetuned to generate SQL queries. This will then be executed on the product's side, to extract relevant production data, with some additional guardrails.
- Another LLM, that's used to formulate an answer to the user query based on relevant data.
- An orchestrator, that manages any back and forth between the Delta's LLMs and our products: this is commonly referred to as chain of thought.
Why RAG?
Foundation LLM models are originally trained on huge datasets containing thousands of different topics: they don't necessarily have knowledge about your specific business and industry. Trying to ask a specific question to such an LLM can either lead to a generic or, even worse, fake answers.
This phenomenon is known as hallucination: LLM is confident in generating false answers based on imaginary facts.
Name
One of the key mechanisms involved in training a neural network is a continuous comparison between actual and predicted values, with the goal of minimizing their difference.
Delta is a Greek letter universally used in science to indicate a difference between two quantities: it seemed appropriate to dedicate our AI engine to this fundamental principle.
There are a couple of additional reasons for choosing this name. First, an uppercase delta visually matches Compiuta's logo quite well: we know have a square and a triangle...you can probably see where this is going. Second, an allusion to one of the most successful Italian rally cars of all time can't hurt, right?