ST-Raptor – An AI table question-answering tool that supports various semi-structured tables

What is ST-Raptor？

ST-Raptor is a tool designed for question answering on semi-structured tables. With just an Excel-format table and a natural language question as input, it can generate precise answers. The tool is capable of handling various semi-structured table layouts, combining vision-language models with a tree construction algorithm, and can flexibly integrate different large language models. ST-Raptor employs a two-stage verification mechanism to ensure reliability of results. It also provides the SSTQA benchmark, which includes 102 tables and 764 questions, for evaluating its performance.

Key Features of ST-Raptor

• Accurate Question Answering: Simply input an Excel table and a natural language question to generate precise answers.

• Support for Diverse Table Layouts: Handles various semi-structured table formats, such as personal information tables, academic records, and financial sheets.

• Multi-format Input: Supports table input from Excel, HTML, Markdown, CSV, and more.

• No Fine-tuning Required: Can be used directly without additional model fine-tuning.

Technical Principles of ST-Raptor

• Vision-Language Models (VLMs): Utilizes VLMs to understand and process visual information in tables.

• Tree Construction Algorithm (HO-Tree): Analyzes and interprets table structures to enhance handling of complex tables.

• Flexible LLM Integration: Supports seamless integration with different large language models, such as DeepSeek-V3 and GPT-4o, to improve QA performance.

• Two-Stage Verification Mechanism: Ensures generated answers are accurate and reliable, reducing the risk of errors.

Project Links

• GitHub Repository: https://github.com/weAIDB/ST-Raptor

Application Scenarios of ST-Raptor

• Enterprise Financial Management: Finance teams can query budget tables to quickly obtain insights for cost control.

• Academic Research Data Management: Researchers can query experimental data tables for specific results, accelerating research progress.

• Human Resources Management: HR professionals can analyze performance tables to assess employee performance and support management decisions.

• Financial Risk Assessment: Analysts can query risk data tables to identify high-risk clients, reducing credit risk.

• Logistics and Supply Chain Management: Managers can query logistics order tables to track inventory and transportation, optimizing the supply chain.