LMQL Is a Programming Language for LLMs (lmql.ai)

LMQL is a domain-specific language embedded in Python that treats prompts as first-class program text and gives developers fine-grained, declarative control over LLM generation. Using an @lmql.query-decorated function, top-level strings become prompts, generation targets are declared as typed variables (e.g., "The answer is [NUM: int]"), and constraints like len(ANSWER) < 120 or STOPS_AT(ANSWER, ".") and hard runtime-enforced clauses via where (e.g., THING in ["Volleyball", ...]) shape outputs. LMQL exposes generated values directly to Python code (the example meaning_of_life() returns NUM == 42), supports expressive control flow and string interpolation for prompt construction, and now adds nested queries to modularize local instructions and reuse prompt components. For the AI/ML community this matters because LMQL turns prompt engineering into programmable, testable, and composable code — improving reliability, format guarantees, and maintainability of LLM-driven systems. Typed outputs and hard constraints reduce parsing and hallucination errors, while nested queries enable encapsulation of subtasks and re-use of context. Practically, that makes building structured question-answering, extraction pipelines, and multi-step reasoning flows safer and more debuggable, bridging the gap between heuristic prompt hacks and robust production-grade LLM orchestration.