Security of the data share is multi-layered.

When you create an order, you are required to add the primary Consumer Point of Contact email address. The data share link generated applies for just that consumer, and only after providing proof of ownership of that email address (via a one-time passcode) are they authenticated into the share.

Once the consumer has authenticated with their email, they can complete the setup of data delivery destinations of their choice, or delegate the destination setup to a colleague via an invite system that restricts invitations to only the email domain of the primary Consumer Point of Contact.

All of the data access methods available are credential-free, meaning that to receive access, the consumer only has to provide non-sensitive details of their cloud account. No static or long-running credentials are ever generated or exchanged in the process.

The Conversational AI interface is linked to the specific data order you create, meaning if you filter or slice data for an order, the consumer would be able to ask questions of and receive answers about only that specific order's accessible data.

We utilize AWS Bedrock for all Large Language Model (LLM) invocations, meaning that your data and the consumer's prompts do not leave the AWS network and are never visible to the model providers for training or any other use.

The LLM has access to your dataset's metadata including table schemas and descriptions as well as other documentation you provide, and it has the ability to generate SQL queries informed by this metadata. Those queries are then run directly on the storage layer data using a data lake query engine (e.g. AWS Athena for AWS S3 storage layer).

Results of these queries can be passed back to the LLM for interpretation, however, these results are normally aggregated and small in size, meaning the LLM would not normally be exposed to the full or even partial raw dataset on which the queries are being run.

By design, the LLM does not have general internet access or search capability to limit the noise that could be introduced to the model in the process of interpreting consumer questions about your data specifically.