Dataherald vs Natural Language SQL

Dataherald is an open-source natural language-to-SQL engine that exposes an API for asking questions about relational data in plain English. Its stated uses include helping business users query warehouses, embedding data Q&A in SaaS products, and creating ChatGPT plug-ins for proprietary data. The project targets developers building natural-language interfaces or contributing to the engine. Self-hosting uses Docker, with MongoDB in a separate container for application data. Listed warehouse connections are PostgreSQL, BigQuery, Databricks, and Snowflake. Included SQL-generation approaches are LangChain SQL Agent, LangChain SQL Chain, LlamaIndex SQL Generator, and Dataherald SQL Agent. A context store can include business and database details, validated question-and-SQL examples, and table information; active learning from usage is also described. An evaluator assigns confidence scores to generated SQL. The modular architecture allows implementations to be changed through .env configuration. Startup requires an OpenAI API key, organization ID, and encryption key. Connection data is encrypted before Mongo storage. The repository uses Apache-2.0 licensing.

Natural Language SQL turns plain-English questions about databases into SQL, validates and runs the query, then streams back results. It connects to PostgreSQL and MySQL, including multiple databases of the same type in parallel. Hybrid retrieval combines BM25 and sentence-transformer vectors to select schema context for a question. Users can review and edit generated SQL before execution, while the interface shows progress through query stages. Results can be sorted, paginated at 50 rows per page, downloaded as CSV or JSON, or copied as tab-separated text. SELECT queries receive a default limit, explicit result counts are capped at 1,000 rows, and execution times out after 30 seconds by default. INSERT, UPDATE and DELETE are allowed by default, with warning banners; the project recommends a read-only database user to block writes. Ollama runs locally by default, keeping questions, schema and generated SQL on the user’s machine. OpenAI, Gemini and Groq are optional cloud providers and receive the question, relevant schema portion and SQL context. The project is free under GPL-3.0. Its recommended Docker setup needs Docker 20.10 or later with Compose v2 and about 10 GB of free disk space.