Relta provides a framework for integrating SQL databases with AI assistants by generating a semantic layer that ensures accurate query handling and data privacy through sandboxed user-specific databases. This technology allows developers to implement natural language interfaces in under two hours while safeguarding against data leakage.
Funding
Funding not disclosed
Founders
Product
Problem
Integrating SQL databases with AI assistants poses challenges in ensuring accurate query handling and preventing data leakage, especially when dealing with sensitive user data. Existing methods often lack the necessary guardrails to guarantee the AI assistant's reliability and data privacy.
Solution
Relta provides a framework that enables developers to integrate SQL databases with AI assistants while maintaining high accuracy and robust data privacy. It achieves this by generating a semantic layer that acts as a guardrail, ensuring user queries are matched to defined metrics and dimensions. Furthermore, Relta creates sandboxed, per-user databases, isolating data and preventing the execution of LLM-generated SQL against the production database, thus mitigating the risk of data leakage. This approach allows for the implementation of natural language interfaces with enhanced security and reliability.
Target Audience
Relta's primary customers are SaaS product developers looking to integrate natural language interfaces into their applications while ensuring data accuracy and privacy when querying SQL databases.
Features
- Semantic layer generation based on sample questions and database schema, defining metrics, dimensions, and measures.
- Sandboxed, per-user databases to prevent data leakage and ensure data privacy.
- Support for Postgres, MySQL, Parquet, and CSV data sources.
- Automatic refinement suggestions for the semantic layer based on user feedback, with automated pull request generation.
- Integrations with LLM frameworks such as Assistants API and LangGraph Cloud.
- Python library for direct integration into existing codebases.