OpenSearch-SQL

OpenSearch-SQL is a framework aimed at converting natural language into SQL queries, designed to reduce the requirement for users' database professional skills. It consists of modules such as preprocessing, extraction, generation, optimization, and alignment, requires no additional training, and can be used in conjunction with models like GPT, DeepSeek, and Gemini.

Main features and innovations:

• Text-to-SQL task: Allows users to ask questions in natural language, and the system automatically generates corresponding SQL queries.
• No additional training required: This framework does not require additional training and can directly use pre-trained large language models (such as GPT).
• Self-learning CoT enhancement: First introduces self-learning chain of thought (CoT) enhancement method, expanding Few-shot format from Query-SQL pairs to Query-CoT-SQL pairs, significantly improving model performance.
• Structured CoT method: Specifically designs a structured CoT method for Text-to-SQL tasks and creates an intermediate language similar to SQL to optimize SQL generation.
• Alignment method: First introduces alignment methods to align the input and output of Agents to alleviate model hallucination problems. This includes:
  • Different styles required by different datasets
  • Hallucinations in information transfer between Agents
  • Basic logic issues in generated SQL
• Excellent performance: Achieves leading position on the BIRD leaderboard.

Usage:

1. Installation: Use pip install -r requirements.txt to install required dependencies.
2. Data processing: Use DAIL-SQL method to generate Few-shot examples, or directly use the provided scripts to generate questions.
3. Run: Execute sh run/run_main.sh to run the main program.

Summary:

OpenSearch-SQL is a powerful Text-to-SQL framework that achieves high performance and ease of use through its unique architectural design and innovative methods, helping users extract information from databases more conveniently. Without requiring additional training, it leverages large language models, enhancing the accuracy and reliability of SQL query generation through chain of thought and alignment methods.