OmniTQA: A Cost-Aware Processing Framework for Hybrid Query of Structured and Unstructured Data

OmniTQA addresses the practical pain points of querying enterprise hybrid data (where structured fields and unstructured text coexist). It elevates semantic reasoning to a first-class query operator, dynamically routes tasks via a dual-engine architecture, and combines data-aware planning and operator-aware batching to achieve dual improvements in accuracy and cost efficiency for scenarios like complex queries and large-scale tables.

In enterprise databases, structured fields (e.g., customer ID, order amount) and unstructured text (e.g., product descriptions, customer service records) often coexist. Traditional Text-to-SQL and table question-answering systems struggle to handle cross-modal reasoning requirements. For example, when a user asks "Products that mention 'eco-friendly materials' in their descriptions and have a return rate below 5% in the past three months", existing methods cannot effectively integrate structured conditions with unstructured text understanding.

The breakthrough of OmniTQA lies in treating semantic reasoning as a "first-class query operator", on par with classic relational operators (selection, projection, etc.), together forming an executable DAG. This design allows the query optimizer to globally optimize the execution plan and provides a unified semantic foundation for hybrid queries.

Fusion of Semantic and Relational Operators

LLM semantic operations are encapsulated as standard query operators, outputting data structures compliant with relational algebra specifications, which can be freely combined with relational operators.

Data-Aware Planning

Minimizes LLM processing load through atomic query decomposition and operator reordering, intelligently offloading structured and semantic tasks.

Dual-Engine Execution

The relational database engine handles structured operations, while the LLM module is responsible for semantic reasoning, dynamically routing tasks; operator-aware batching merges similar LLM requests to improve throughput.

OmniTQA significantly outperforms existing symbolic, semantic, and hybrid baselines in diverse benchmark tests, especially excelling in scenarios like complex queries, large-scale tables, and multi-relation schemas. Meanwhile, by reducing LLM calls and optimizing batching, it drastically lowers processing costs while ensuring accuracy.

OmniTQA solves hybrid query pain points in scenarios like customer relationship management and e-commerce search (e.g., the query "Phones with reviews mentioning 'high cost-performance' and priced between 500-1000 yuan" in e-commerce). It represents an important direction for the integration of databases and LLMs, and its progressive evolution path facilitates enterprise technology upgrades.

In the future, OmniTQA can support more unstructured data types (images, audio), enhance the reasoning capability of semantic operators, and explore more aggressive query optimization strategies. Such cost-aware frameworks will become key for enterprises to handle intelligent queries of large-scale hybrid data.