Large language models have dominated the conversation in AI for the last few years - ChatGPT, Claude, and their competitors have made "the model understands language" a familiar idea. But quietly, alongside that wave, a different kind of model has been taking shape: one trained not on human writing, but on the other great untapped data source inside every organization - the enterprise database. These are being called Large Database Models, or LDMs.
The Core Idea
LLMs are trained on books, articles, code, and the open web - anything written in human language. They learn the meaning of words: which ones tend to appear together, what concepts relate to each other, how ideas connect. LDMs do the same thing, but for data records. Instead of training on sentences, they train on the rows and columns already sitting in a company's databases - every transaction, customer profile, purchase, claim, and support ticket. And just as an LLM learns what a word "means" by its context, an LDM learns what a value means - a customer's region, spending pattern, or product category - by how it co-occurs with other values across millions of records. The result is a model that understands your database the way an LLM understands language: not just as rows to be filtered, but as records with meaning and relationships to each other.
Why This Matters: Semantic Queries
Here's the practical difference it unlocks. A traditional database query has to be written as an exact, rigid filter - you tell it precisely which conditions to match:
"List all customers who live in California, are over 40, and have spent at least $2,000."
That's useful, but it only works if you already know exactly which conditions matter, and it can't find similar records - only exact matches on the rules you've written. An LDM lets you skip that rule-writing entirely and instead ask a fundamentally different kind of question, one based on meaning and resemblance rather than explicit filters:
"List all customers most similar to Jane Doe."
"List all products most similar to this one, based on how customers actually use them."
"Which transactions deviate from this account's normal pattern?"
"Which past applicants does this new applicant most resemble, and what happened to them?"
"Which support tickets are really the same underlying issue, even if worded differently?"
These are semantic queries - and they don't require you to specify which fields matter or how to weight them, or to write any SQL to define "similar." The model has already learned that from the data itself.
How Is This Different From Traditional Machine Learning?
It's worth being precise here, because LDMs aren't really trained to output a single score for one specific outcome - they're closer to a search and matching layer. A traditional ML model is built to answer one narrow question (will this customer churn? is this transaction fraudulent?) and needs to be retrained whenever the business question changes.
An LDM, by contrast, doesn't answer just one fixed question - it finds similar records, and lets you infer the answer from what happened to those similar records historically. Instead of building a separate model for every question, you're reusing the same underlying representation of your data to answer many different questions - fraud risk, churn risk, recommendation relevance, and more - all through the same mechanism: nearest-neighbor similarity.
This also makes results more explainable. Instead of a black-box score, you can point to the actual similar records driving a decision: "this application was approved because it closely resembles a group of past applicants who repaid on time."
A Real-World Signal This Is Already Working
This isn't just a research concept. Database vendors have already begun shipping LDM-style capability as a built-in feature, and early adopters - including large insurance companies using it to gauge the likelihood that a sales quote will convert - have reported measurable, fast improvements in conversion rates within months of deployment, without needing a team of data scientists to build and maintain a custom model.
That's the real promise of LDMs: capability that used to require a dedicated machine learning team is starting to become something database and analytics teams can use directly.
What's Next
LDMs are still early - nowhere near as mainstream as LLMs yet - but the direction is clear: just as LLMs made the world's written knowledge queryable in plain language, LDMs are starting to do the same for the world's structured business data.
In the next blog, we'll walk through ten concrete business use cases where this kind of similarity search over your own data can create real value - from fraud detection and credit scoring to demand forecasting and personalized recommendations.