What Is LMQL and What Is the lmql.ai Website For?

LMQL (Language Model Query Language) is a programming language for interacting with large language models, and lmql.ai is its official website. It is meant for developers who want LLM prompting to be robust and modular rather than a pile of loosely managed strings. The site describes LMQL as using "types, templates, constraints and an optimizing runtime," and it was created by the SRI Lab at ETH Zurich along with contributors. If you write Python and want more control over model output, LMQL is the relevant tool; if you just want a chat interface, it is more than you need.

What problem LMQL solves

Plain prompting usually means sending text and hoping the model answers in the format you need. LMQL treats prompting as a program: prompts are written as top-level strings, generation happens through variables, and constraints are enforced by the runtime. The site's own example asks for the answer to life, the universe and everything, constrains the answer to fewer than 120 characters and to stop at a period, then prints the result directly. A second part uses a typed variable ([NUM: int]) so the output format is guaranteed. The point is that output shape and stopping conditions become part of the query, not something you clean up afterward.

Constraints and types

Constraints such as length limits and stop conditions are checked by the runtime while the model generates, and typed variables like integers give you a guaranteed format. This is the main difference from ordinary prompt engineering: the program, not the reader, enforces the rules.

Python integration

LMQL queries are Python functions. You decorate a function with @lmql.query, write the query inside a docstring, and call it like any other function. Results are accessible as variables, so you can branch, loop and reuse them in normal Python code.

What you can build with it

The site lists several capabilities that show the intended scope:

• Constrained LLMs — enforce output rules during generation.
• Types and regex — control the exact shape of generated values.
• Multi-part prompts — compose prompts from several sections.
• Nested queries — the site highlights this as a newer feature, enabling modular local instructions and reuse of prompt components.
• Python support — prompt construction via Python control flow and string interpolation, including loops for repeated prompt sections.
• Measure distributions, meta prompting, tool augmentation, chatbots — broader use cases listed on the site.

A concrete example from the site is a packing list: a loop generates four items, and a where clause constrains each item to a fixed set such as "Volleyball," "Sunscreen" or "Bathing." That is a small illustration of the general pattern — the runtime enforces the allowed values instead of you validating them afterward.

Backend portability

LMQL code is portable across several backends, and the site says you can switch between them with a single line of code. The backends named are llama.cpp, OpenAI and Hugging Face Transformers. This matters if you want to prototype against a hosted model and later run locally, or compare behavior across models without rewriting prompts.

What the lmql.ai website offers

The site is the project's hub rather than a product page. Its main navigation includes:

The homepage also links to a developer survey for shaping the next major version, and notes the project is open to contributions. There is no pricing information on the page, so nothing here should be read as confirming the tool is free or requires no account — check the Docs and GitHub for current terms.

Who should use it

LMQL fits you if you are comfortable in Python, need reliable output formats from an LLM, and want constraints enforced during generation rather than patched afterward. It also fits teams that want to move prompts between local and hosted backends. If your task is a one-off question to a chatbot, the language and its runtime add overhead you will not use. The fastest way to judge it is to open the Playground, run the example query, then read the Docs for the constraint syntax before committing to it in a project.