Skip to main content

LLM & data privacy

Zingle's agents use a large language model (LLM) to reason about your dbt project and generate SQL, YAML, tests, and documentation. This page describes what is sent to the model and how.

Which model provider​

Zingle runs on Anthropic's Claude models. Inference can be served through:

  • AWS Bedrock - inference runs within AWS. Amazon Bedrock does not use your prompts or completions to train foundation models.
  • The model-provider API - a direct API to the model provider.

What is sent to the model​

To do useful work, the agent shares relevant context with the model, which can include:

  • Project code and metadata - the dbt models, columns, and files relevant to your request.
  • Tool results - the outputs of the tools the agent runs, which can include rows returned by warehouse queries (capped per query) when a query is part of the task.

In other words, when a task involves querying your warehouse, the query results can be included in the model context so the agent can reason over them.

Minimizing exposure

The agent is designed to retrieve only the context relevant to the task rather than your entire project or warehouse. Warehouse results are always row-capped.

What is not sent​

  • Credentials and secrets are not part of the prompt content sent to the model.

Training​

Your data is not used to train foundation models when inference is served through AWS Bedrock.