Skip to main content

Key Concepts

This glossary defines key terms as they are used throughout the nilAI documentation and system.

Attestation​

A cryptographic proof mechanism that verifies the integrity of the TEE environment. nilAI provides two layers of attestation:

  • Environment Attestation: Verifies the GPU environment and service integrity through a verifying key
  • Response Signatures: Cryptographic signatures included with every chat completion response, proving the response originated from the attested environment

Chunks​

Segments of data created when files are processed for nilRAG. When a data owner uploads files:

  • Files are split into multiple chunks of data
  • Each chunk has a corresponding embedding for similarity search
  • Chunks are blinded before being uploaded to Private Storage
  • During retrieval, chunks provide the actual content for LLM context

Embeddings​

Numerical vector representations of text used for similarity search in nilRAG:

  • Generated from chunks when data owners upload files
  • Used to match client queries against stored content
  • Stored in secret-shared form in Private Storage
  • Enable privacy-preserving similarity comparisons without exposing raw data

NVIDIA Confidential Computing​

The hardware security technology that enables TEE-based AI inference in nilAI. NVIDIA Confidential Computing provides GPU-level isolation and encryption, ensuring that AI model processing remains private even from the infrastructure provider.

OpenAI-Compatible API​

The interface standard used by nilAI for AI inference. Existing applications using OpenAI-style code can migrate to private LLMs by changing the endpoint URL while keeping the same code structure and API calls.

Signed Responses​

Cryptographic signatures included with every chat completion response from nilAI. These signatures allow verification that:

  • The response came from an attested TEE environment
  • The response has not been tampered with
  • The processing occurred in a verified secure enclave