Wanted to share a project I have been working. Data Connector is an open source tool for interacting with Databases (Postgres, MySQL, SQLite) using natural language. Prompts are handled using a ReAct[1] agent which has access to the schema and a tool for executing queries.

It also has experimental[2] support for calling APIs (OpenAPI, GraphQL) which allows for queries that can span multiple systems (Find this user's details in Clerk, etc).

We have mostly been using it for managing our local development environment, but it has configurable controls for running in other environments.

- Privacy Mode which never sends query results back to the LLM context, instead the result is returned to the user as JSON

- Approval Mode which requires manual approval of each query before it is executed

Data Connector runs within a private subnet, connects to the database and polls for new queries. There is no requirement to open traffic to the database or share credentials / connection strings.

Data Connector is built using another project we are working on Inferable (also open source) which is a runtime for building and observing LLM-based agents.

Would love your thoughts!

[1] https://arxiv.org/abs/2210.03629

[2] We have been testing these with various schemas and are currently working on better handling for large OpenAPI specs, etc.