-
· Fundamentals · Choosing an approach
Should an Assistant Answer from Database Rows or Documents?
Route assistant questions by the fact they need: rows for counts, documents for rules, both for combined ones, with the route and sources shown to the reader.
-
· 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.
-
· Fundamentals
\"Could not finish the message because max_tokens\": reasoning models and tiny token budgets
Why a reasoning model meets a small max_tokens with an HTTP 400 or an empty reply, why max_output_tokens has a minimum of 16, and what nlqueries-core 0.4 does.
-
· Fundamentals
Semantic caching for text-to-SQL: cheaper, faster answers to repeat questions
How a semantic cache for LLM-generated SQL safely reuses answers to repeat questions in nlqueries-core 0.4, what entries are bound to, and what expires them.
-
· Fundamentals
Text-to-SQL, explained: how a question becomes a query, and where it goes wrong
How text-to-SQL works end to end: schema extraction, the knowledge base, generation, validation, execution and the one correction after a database error.
-
· Fundamentals
Text to SQL is not the hard part
Getting a language model to emit SQL is close to solved. Getting an organisation to trust the answer is not, and that needs different solutions.
-
· 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.
No posts match your search. Search all posts