SEMA

SEMA is a multimodal semantic data system for querying and transforming heterogeneous data with LLM-powered semantic operators.

Problem

AI data pipelines need to process text, tables, images, video, and semi-structured records in one workflow. Traditional query engines are strong at relational operators, but they do not expose semantic interpretation as an optimizable first-class database operator.

Core idea

SEMA embeds LLM semantic operators into database query plans. Through SemaSQL, users express semantic querying and ETL in a SQL-like form; the engine then optimizes execution with operator reordering, semantic fusion, prompt batching, and adaptive execution.

My role

System design and architecture for the semantic query-engine direction, including semantic operator abstraction, query-plan integration, and execution optimization around token cost and latency.

Evidence

• Paper
• First-class LLM semantic operators inside DuckDB-style query plans.
• Adaptive execution techniques including operator reordering, semantic fusion, and prompt batching.

Impact

SEMA shows that semantic ETL and ad-hoc semantic querying can be managed inside a query engine. Reported semantic-query speedups are 2–10×, while preserving a database-style interface for heterogeneous AI data processing.