The Problem with Traditional Data Analysis

Most enterprise data—estimated at 99%—remains locked within relational databases, inaccessible to Large Language Models (LLMs). Traditional analytical workflows require data scientists to manually select specific columns for SQL queries, which is rigid, slow, and expensive. Furthermore, moving data out of secure, regulated environments for analysis increases security risks and consumes significant IT budgets (reportedly 32% to 40%).

How Large Database Models (LDMs) Work

LDMs bridge the gap between AI and structured data by training models directly on database tables. The process follows five key steps:

  1. Classification: Select a table and classify columns as categorical (discrete values) or numeric (continuous values).
  2. Tokenization and Binning: Categorical values are treated as tokens. Numeric values are binned into clusters (e.g., ages 37 and 38 become the same bucket ID) to ensure the model treats numerically close values as semantically similar.
  3. Row-to-Sentence Conversion: Each row is converted into an unordered "bag of words" where every token (column name + value) has an equal relationship to others in the row.
  4. Training: A self-supervised neural network learns vector representations for each token. Values that appear in similar contexts (rows) are mapped to nearby points in vector space.
  5. SQL Integration: The trained model is loaded back into the database, allowing users to perform semantic queries using standard SQL.

Practical Applications and Benefits

LDMs allow users to perform complex tasks like similarity searches, anomaly detection, and clustering without needing a data scientist to build an external pipeline. Because the model runs where the data lives, it maintains security and governance standards.

Industry use cases include:

  • Insurance: Predicting successful quotes by retrieving similar past contracts.
  • Fraud Detection: Flagging transactions that deviate from established patterns.
  • Retail/Food: Identifying product similarities (e.g., finding nutritional alternatives) based on database attributes.

IBM introduced this technology via SQL Data Insights for DB2, which has since evolved into SQL Data Insights Pro, adding support for unstructured text and incremental model updates.