Oracle Cloud Infrastructure Documentation

Data Handling in Generative AI

Find answers to common questions about how OCI Generative AI uses, retains, and protects your data.

Select your use case:

Imported Models

Are my inference requests sent to Hugging Face?

Your inference requests run on the dedicated AI cluster that hosts your imported model. Importing a model from Hugging Face retrieves model artifacts for deployment. It doesn't send inference requests to Hugging Face. See Compatible Models for Import.

Are my prompts and responses retained or used for model training?

The service doesn't retain your prompts and responses as part of model inference. Model providers don't use them for their own model training, model improvement, or other independent purposes.

You can choose to save responses, conversations, files, or memory for your application through OCI features. For how you manage this content, see Stored Responses, Conversations, Files, and Memory.

Do I need to provide training data when I import a model?

The import workflow requires model files, such as weights, configuration, and tokenizer files. It doesn't require the dataset used to train the model.

You can import a compatible model that you fine-tuned outside the service. OCI Generative AI doesn't fine-tune imported models. See Importing Fine-Tuned Models.

Can I delete my imported model?

You can delete the imported model resource. You manage this resource separately from individual inference requests. See Deleting an Imported Model.

Pretrained and Fine-Tuned Models on Dedicated AI Clusters

Are my prompts sent to an external model provider?

Your prompts are processed by the model on your dedicated AI cluster and aren't sent to an external provider for inference. The cluster is dedicated to your models and isn't shared with other customers. See Dedicated AI Clusters.

Are my prompts and responses retained or used for model training?

The service doesn't retain your prompts and responses as part of model inference. Model providers don't use them for their own model training, model improvement, or other independent purposes.

You can choose to save responses, conversations, files, or memory for your application through OCI features. For how you manage this content, see Stored Responses, Conversations, Files, and Memory.

Which model can I fine-tune in OCI Generative AI?

Currently, the only model available for fine-tuning in the service is the standard Meta Llama 3.3 (70B) model (meta.llama-3.3-70b-instruct). Fine-tuning is available in supported commercial (OC1) regions. The dynamic FP8 variant isn't available for fine-tuning.

Does the service keep my training data after fine-tuning?

The service deletes training data from the fine-tuning cluster when the training job completes. Your source dataset remains in your Object Storage bucket, where you manage it. You also manage the resulting custom model as a separate resource.

Is my training data or custom model shared with model providers?

The service doesn't share your fine-tuning training data or fine-tuned custom model with third-party model providers. It uses your training data only to build your custom model, and doesn't use it to improve general OCI Generative AI models. See Fine-Tuning the Base Models in Generative AI.

Is my training data encrypted?

Your fine-tuning training data is encrypted at rest and in transit. For encryption and key management details, see Encryption.

On-Demand Models

How do I know whether my model makes calls to an external provider?

Find your model and the OCI region that serves your request in Models by Region. If you use a routing profile, check its target OCI regions. The External Calls notes identify models, including Google Gemini and xAI Grok, whose requests are sent to an external provider for processing.

Google Gemini and xAI Grok

Are my prompts sent to Google or xAI?

When you call these models through OCI Generative AI, the service sends your request content to Google or xAI for processing on the provider's platform and returns the generated response to you.

Can Google or xAI retain my prompts and responses or use them for model training?

Oracle has contractual Zero Data Retention agreements with Google and xAI for models accessed through OCI Generative AI. Under these agreements, the providers use customer prompts and responses only to process requests and deliver responses. They don't retain this content after processing the requests or use it for their own model training, model improvement, or other independent purposes.

Does OCI Generative AI retain my prompts and responses?

The service doesn't retain your prompts and responses as part of model inference.

You can choose to save responses, conversations, files, or memory for your application through OCI features. For how you manage this content, see Stored Responses, Conversations, Files, and Memory.

Where do Google and xAI process my requests?

The processing location depends on the model and the OCI region that serves your request. If you use a routing profile, use the notes for the target OCI region that serves the request. This can differ from the region where you create or call the profile. See the model-specific routing and processing locations in Notes for External Calls.

Cohere, Meta Llama, and OpenAI gpt-oss

Are my inference requests sent to an external model provider?

Cohere, Meta Llama, and OpenAI gpt-oss models offered in on-demand mode are hosted by OCI Generative AI. The service processes your requests without sending them to an external provider for inference. For supported models and regions, see Models by Region.

Are my prompts and responses retained or used for model training?

The service doesn't retain your prompts and responses as part of model inference. Model providers don't use them for their own model training, model improvement, or other independent purposes.

You can choose to save responses, conversations, files, or memory for your application through OCI features. For how you manage this content, see Stored Responses, Conversations, Files, and Memory.

On-Demand Models Through Routing Profiles

Which data-handling rules apply when I call an on-demand model from a region with no local model processing?

The on-demand data-handling rules for the model selected in your routing profile apply. The profile routes your request to one of the target OCI regions that you select. For models that make calls to an external provider, review the External Calls notes for that model and the target OCI region that serves the request.

For example, from Australia East (Sydney), you can create a routing profile with Japan Central (Osaka) as its only target OCI region and select one of these models:

For supported models and target OCI regions, see Models by Region. If you select multiple target OCI regions, check the External Calls notes for each region. For routing instructions, see Regions with no Local Model Processing.

Stored Responses, Conversations, Files, and Memory

What happens when I choose to save responses, conversations, files, or memory?

You create an Enterprise AI Agents project in a compartment in your OCI tenancy and choose the features your application needs. When you use features that save responses or conversations, upload files, or enable long-term memory, you are choosing to retain content for your application. OCI saves this content as part of your project so that your application can retrieve or reuse it.

You configure how long responses and conversations are retained. You choose which files to upload and manage through the Files API. If you enable long-term memory, the service saves information from conversations for use in future interactions within the project. See Creating a Project, Data Retention, Agent Memory, and Files API.

Can I save content in my project when I use Google Gemini or xAI Grok?

You can choose to save content for your application through OCI project features. Oracle's Zero Data Retention agreements with Google and xAI govern the content those providers process when you call their models. Content you choose to save in OCI is managed as your project data and follows the applicable OCI storage and retention settings.

Encryption

How is fine-tuning training data protected?

Training data is encrypted at rest and in transit. The service deletes the training data from the fine-tuning cluster as soon as the training job completes.

Training data for fine-tuning a model is always double encrypted, by default, by Oracle-managed AES-256 encryption and optionally by customer-managed private keys through the OCI Vault service. You can delete the data at any time.

Oracle encrypts data in motion with TLS 1.2.