This page covers information that is more relevant to self-hosting. For a more comprehensive overview of how data is processed in Onyx, check out Data Flows and Data Storage

​
Complete Data Isolation

All data processing occurs within your infrastructure, ensuring complete control over your data. No sensitive data leaves your network except for the specific communications you configure. Note, anonymous telemetry is enabled by default but can be turned off at deployment time.

​
Local Model Execution

  • Data Processing: Document parsing, chunking, and metadata extraction done on-premises
  • Embedding/Reranking Models: All document embeddings happens locally
  • LLM Integration: Bring your own LLM - no data sent to external services unless explicitly configured
  • Custom Models: Other models used by Onyx are trained to be runnable on both CPU or GPU.

​
Data Flow Security

​
Inbound Data

  • Connector Data: Document content and metadata updates get pulled in from configured external applications
  • Model Downloads: If you configure a non-default embedding or reranking model, it will need to be downloaded over the network (it still runs locally).
  • User Interactions: Chat messages and actions taken in the UI.
  • User Uploads: Documents uploaded in the Chat interface by the user get processed and stored locally.

​
Outbound Data

  • LLM Inference: Queries, chat history, and documents sent to the LLM for processing.
  • Actions: Calls to external APIs configured by the admin users in your Onyx deployment.
  • Web Search: Queries passed to a search provider and scraper of choice.
You can also configure external document processing services, embedding/reranking APIs, image captioning models, etc. but these are purely optional and have built-in equivalents.