Semantic query processing

Semantic query processing engines are a new class of analytical data processing system. They expand SQL with semantic operators, performing operations that require semantic understanding of data. For instance, whereas traditional database systems can only count the nnumber of reviews containing specific keywords, semantic engines can count the number of reviews that are positive about the product.

Such operators are enabled by large language models (LLMs) that perform various tasks, based on a natural language description of the task alone. However, applying LLMs to data is extremely expensive, shifting the performance bottleneck in query processing entirely to semantic operators.

We design novel systems and algorithms that make semantic query processing more cost-efficient. This includes, for instance, the ThalamusDB system, leveraging approximate query processing techniques in the context of semantic analysis. It also includes novel join operator implementations, minimizing the number of tokens consumed when processing semantic joins. Finlly, it includes techniques, such as SpareLLM, that select LLMs realizing optimal tradeoffs between cost and quality.