Using LLMs for Data Analysis and SQL Query Generation
(Run this example in Google Colab here (opens in a new tab))
Large language models (LLMs) like Nous-Hermes-Llama2-13B and WizardCoder have demonstrated impressive capabilities for understanding natural language and generating SQL. We can leverage these skills for data analysis by having them automatically generate SQL queries against known database structures.
Unlike code generation interfaces that attempt to produce executable code from scratch, our approach focuses strictly on generating industry-standard SQL from plain English questions. This provides two major benefits:
SQL is a well-established language supported across environments, avoiding the need to execute less secure auto-generated code.
Mapping natural language questions to SQL over known schemas is more robust than attempting to generate arbitrary code for unfamiliar data structures.
By combining language model understanding of questions with a defined database schema, the system can translate simple natural language queries into precise SQL for fast and reliable data analysis. This makes surfacing insights more accessible compared to manual SQL writing or hopelessly broad code generation.
Prediction Guard provides access to such state-of-the-art models that maintain strong capabilities while including safety measures to mitigate potential harms. We'll walk through an example of using these LLMs for data analysis on sample jobs dataset from Kaggle.