What is ST-Raptor
ST-Raptor is a tool for question answering over semi-structured tables. It only requires an Excel-format table and a natural language question as input, and can produce precise answers. The tool can handle a variety of semi-structured table layouts, combines vision-language models and tree construction algorithms, and can flexibly integrate different large language models. ST-Raptor has a two-stage verification mechanism to ensure reliable results. ST-Raptor provides the SSTQA benchmark, which contains 102 tables and 764 questions, for evaluating its performance.
ST-Raptor's main features
-
Precise question answering: It only requires an Excel-format table and a natural language question as input, and can generate precise answers.
-
Diverse table support: Supports a variety of semi-structured table layouts, such as personal information forms, academic tables, financial tables, etc.
-
Multiple input formats: Supports table input from multiple formats such as Excel, HTML, Markdown, CSV, etc.
-
No fine-tuning required: It can be used without additional fine-tuning of the model.
ST-Raptor's technical principles
-
Vision-language model (VLM): Combined with vision-language models, it can understand and process visual information in tables.
-
Tree construction algorithm (HO-Tree): Through the tree construction algorithm, it analyzes and understands table structures, improving its ability to handle complex tables.
-
Flexible integration of large language models (LLM): Supports flexible integration of different large language models, such asDeepseek-V3、GPT-4o, etc., to improve question-answering performance.
-
Two-stage verification mechanism: Through the two-stage verification mechanism, it ensures that the generated answers are accurate and reliable, avoiding the production of incorrect answers.
ST-Raptor's project address
- GitHub repository：https://github.com/weAIDB/ST-Raptor
ST-Raptor's application scenarios
-
Corporate financial management: Finance personnel input a budget table and can quickly obtain answers, helping with financial cost control.
-
Academic research data management: Researchers input an experimental data table and can query specific results, accelerating the research process.
-
Human resource management: HR inputs a performance table and can query employee performance, providing support for management decisions.
-
Financial risk assessment: Analysts input a risk data table and can identify high-risk customers, reducing credit risk.
-
Logistics supply chain management: Managers input a logistics order table and can query inventory and transportation status, optimizing the supply chain.