Skip to content

Architecture

Relta is packaged as a Python library. The library exposes methods that are used to:

  1. Connect relational datasources and configure semantic layers from them at set-up

  2. Implement natural language chat sessions during runtime

During set-up, DataSource and SemanticLayer objects are configured and automatically stored in a .relta folder in your repo. These will get copied into each Python process spun up during runtime, allowing each user session to access the latest configuration of connected DataSource and SemanticLayer objects. At runtime Relta will also spin up and store in-process DuckDB instances in the .relta folder which will be used to create a sandboxed replica of that user's data from the underlying datasource.

There is currently no stand-alone server version of Relta. However, the CLI provides convenient interfaces for initial set-up.

We suggest the following implementation:

  1. Configure Relta using the CLI in your backend repository:
    • Define all the DataSource objects that you want to use with Relta.
    • Generate the SemanticLayer objects for each DataSource object using the CLI
    • Write any other changes to Configuration to pyproject.toml
  2. Use the Client object in your backend code to create secure sandboxed databases and implement chat sessions:
    • Fetch relevant DataSource and SemanticLayer objects using get_datasource
    • Create a local, sandboxed database with the user's data using deploy
    • Create a chat for the user with create_chat
    • Refine the SemanticLayer from user feedback using the .feedback method on a Response object, which is returned from chat.prompt.

Architecture

One Relta Client per Python Process

A current limitation of Relta is that there should only be one relta.Client instance per Python process, implying that only one Python process should be used per user.

Relta Server

We are working with select customers to design a server that batches feedback, caches user sessions, and persists chat histories. If you are interested in this design partnership, please reach out to us at this link.

  1. Brain icons created by Umeicon - Flaticon ↩