LMQL (Language Model Query Language) is a specialized programming language designed for interacting with large language models in a structured, controllable way. It combines natural language prompting with programmatic constraints and control flow, giving developers precise control over LLM generation.

Key Concepts

• Query Syntax: LMQL uses a SQL-like syntax where you write prompts as queries with embedded constraints on the generated output.
• Constraints: You can specify rules like "output must be one of [list]", "output length must be < N tokens", or "output must match a regex pattern" — and LMQL enforces these during generation.
• Control Flow: Supports Python-like control flow (if/else, for loops) within prompts, enabling dynamic, branching conversations.
• Scripted Interaction: Multi-turn interactions can be scripted as a single LMQL program rather than managing state manually.

Example Capabilities

• Type Constraints: Force outputs to be valid integers, booleans, or selections from enumerated options.
• Length Control: Limit generation to a specific number of tokens or characters.
• Decoder Control: Specify decoding strategies (beam search, sampling with temperature) per generation step.
• Nested Queries: Compose complex prompts from simpler sub-queries.

Advantages Over Raw Prompting

• Reliability: Constraints guarantee output format compliance, eliminating the need for post-hoc parsing and retry logic.
• Efficiency: Token-level constraint checking can prune invalid tokens before they're generated, saving compute.
• Debugging: LMQL programs are structured and testable, unlike ad-hoc prompt strings.

Integration

LMQL supports multiple backends including OpenAI, HuggingFace Transformers, and llama.cpp. It can be used as a Python library or through its own interactive playground.

LMQL represents the trend toward treating LLM interaction as a programming discipline rather than an art of prompt crafting.