Design and method
Knowledge system for chat-to-data
A structured knowledge base that sits between a company’s database and an AI assistant, so a question in plain English becomes correct SQL.
The pain
AI writes confident queries that are wrong for a specific business, because it does not know what the company’s data means.
AI models write plausible SQL that is wrong for a specific business: the wrong join, the wrong definition of revenue, a column that looks right and is not.
The fix is giving the model the business’s own context in a form it can read reliably.
What was built
One layout per client and database
A new client is added without changing the global structure. Each declares its SQL dialect.
An article per table
Purpose, columns, and the caveats a long-time analyst would know.
Metric definitions
Kept as a human-readable page and a machine-readable twin, with a script that checks the two agree.
A graph an agent can walk
Tables, columns, foreign keys, and which metrics use which columns, so an agent can plan a query.
A worked sample
A healthcare-claims database with four documented tables, two metrics and their lineage. Query logs are designed as one file per day, for audit.
Results
There is no accuracy benchmark for this project yet. The same idea ran in production in my work at a real-estate software company.
How it was built
- Human-readable and machine-readable definitions are checked against each other by script.
- Guidance for the model and a separate guide for the people who use it.
- Query logging designed for audit, secured at the storage layer.
What it means for a business
“Ask your data a question” only works when the AI knows your definitions. This is the method for capturing them once and keeping them correct as the business changes.