Skip to content

Data Storage & The Knowledge Core ​

ANDARTIS is designed with a strict "Zero Data Egress" mandate. All documents, extracted metadata, vector indices, and neural weights are stored locally on your machine.


The Knowledge Base (SQLite) ​

To guarantee high scalability, zero lock contention, and simple backup processes, ANDARTIS uses an Isolated Node Core design.

Rather than storing all user data in a single massive database, each added directory (Intelligence Node) gets its own isolated SQLite database file.

Main Database (main.sqlite) ​

Acts as a registry and orchestrator schema:

  • intelligence_nodes: Registry of directory paths, titles, and config states.
  • capabilities: Definitions, active weights paths, and options for registered brains.
  • projects: Visual group containers for grouping related intelligence nodes under project namespaces.
  • node_capabilities: Pivot table mapping nodes to capabilities and tracking training/activation status.
  • cross_reasoning_groups: Groups of nodes within a project namespace configured for cross-node reasoning.

Node Database (node_{id}/core.sqlite) ​

Stores the raw and structured representations of your documents:

  • documents: Reference metadata, unique file paths, hashes, and document-level metadata (containing schema extractions per file).
  • document_chunks: Fragmented sections of document content (plain text) along with chunk-level metadata for fine-grained retrieval.

Neural Weights (.npz / .safetensors) ​

While textual metadata resides in SQLite, the weights representing trained models are stored in the local file system.

File Structure per Node ​

Each intelligence node has a corresponding directory in the storage folder:

storage/models/node_{id}/
├── weights.npz       # Learned parameters of MLX models
├── index.npz         # High-speed index for semantic navigation
└── metadata.json     # Configuration file (vocab, tokenizers)
  • weights.npz: Stores custom layers adjusted during training.
  • index.npz: An optimized array for nearest-neighbor search, allowing sub-millisecond semantic matching.

NativePHP Directory Mapping ​

When packaged for production, directories follow standard operating system application support conventions:

macOS Application Storage ​

~/Library/Application Support/andartis/

Windows Application Storage ​

%APPDATA%\andartis\

Privacy & Security ​

IMPORTANT

ANDARTIS does not provide an internal network layer or cloud sync tool. All operations are local-only.

  1. Physical Isolation: ANDARTIS reads documents from folders you choose. It does not copy, upload, or move them.
  2. Encryption-at-Rest Compatibility: Because the backend uses standardized files (SQLite, safetensors, JSON), you can secure your intelligence nodes using native system-level utilities like macOS FileVault or Windows BitLocker.
  3. No Telemetry: There are no tracking scripts, telemetry calls, or remote metrics. The app functions entirely air-gapped.

Released in Alpha