-
· Accuracy & validation
How to Compare AST-Validated NL2SQL with Raw LLM SQL: Measuring Schema Validity, Policy Enforcement, and Answer Correctness
Measure schema validity, SQL policy enforcement and answer correctness separately when comparing validated NL2SQL with raw LLM SQL.
-
· Fundamentals · Accuracy & validation
A Reliable NL2SQL Architecture: Separate Generation, Validation, Execution, and Explanation
Why "revenue was $0" happens in NL2SQL apps, and how separate generation, validation, execution and explanation stages help prevent it.
-
· Accuracy & validation · Choosing an approach
Schema Prompting vs. Schema Retrieval: Choosing an NL2SQL Context Strategy
Full-schema prompting or retrieved context? How to measure table recall, join closure and wrong joins before choosing an NL2SQL context strategy.
-
· Accuracy & validation
When the database says no: self-correcting SQL from database errors
How LLM SQL error correction works in nlqueries-core 0.4: one retry after a database error, the error framed as data, and the errors never retried.
-
· Accuracy & validation
From thumbs-up to better SQL: turning user feedback into verified examples
How a text-to-SQL feedback loop works in nlqueries-core 0.4: ratings and corrections become examples the model sees, after checks on who gave them.
-
· Accuracy & validation
How to build a text-to-SQL knowledge base from query history
Query history shows an LLM the joins and filters your team actually uses. How process-history and export-kb build it into the text-to-SQL knowledge base.
-
· Accuracy & validation · Security & governance
Why we validate SQL before it runs
The difference between a query that parses, a query that executes, and a query that answers the question. Three separate checks, and only one of them is free.
-
· Accuracy & validation
Your query history knows what your schema does not
Schema introspection tells you a column is called status and holds a varchar. Years of queries tell you which joins and filters anyone actually uses.
-
· Fundamentals · Accuracy & validation
The join problem
Single-table questions are easy. Most real questions are not, and when a join goes wrong the result is a plausible number rather than an error.
-
· Accuracy & validation
How to evaluate a text-to-SQL system on your own data
Public leaderboard scores do not predict how a system will do on your warehouse. Here is the evaluation we would build if we were buying rather than selling.
No posts match your search. Search all posts