LMQL is a query language designed for large language models (LLMs) that enhances the interaction with these models by providing control, flexibility, and customization. Key Features: - Constraints: Users can specify high-level logical constraints over the language model output, which are automatically converted into token-level prediction masks. - Debugging: LMQL includes a Playground IDE for query development, allowing users to inspect the interpreter state, validation result, and model results. - Efficiency: LMQL uses novel partial evaluation semantics to evaluate and control the language model decoding process on a token level, leading to significant efficiency gains. - Frontend/Backend Separation: LMQL provides a high-level frontend to interact with language models, abstracting over model-specific implementation details and allowing for easy development and quick prototyping. - Retrieval: LMQL supports advanced text processing and retrieval of information from language models.