Memelang: An Axial Grammar for LLM-Generated Vector-Relational Queries (arxiv.org)

A recent paper introduces Memelang, a novel axial grammar designed to enhance the generation of vector-relational queries by large language models (LLMs). This innovation emphasizes a compact domain-specific language (DSL) that can be directly emitted and deterministically parsed, allowing LLMs to create structured SQL queries with greater efficiency. Utilizing linear token sequences and a unique method of placing rank-specific separator tokens, Memelang organizes data into an n-dimensional grid without the clutter of traditional parentheses or complex syntax. The significance of Memelang lies in its ability to streamline query generation, paving the way for improved interaction between LLMs and database systems. Key features include coordinate-stable references, parse-time variable binding, and mechanisms for implicit context carry-forward, which collectively minimize redundancy in generated queries. Additionally, the grammar supports essential functions like grouping, aggregation, and ordering via inline tags, enabling seamless derived execution plans in a single pass. The paper provides practical tools, including a lexer/parser and a compiler for PostgreSQL SQL, which could greatly benefit developers and researchers in AI/ML by simplifying how queries are constructed and executed.