Data Storage & Embedded Databases

In compliance with the project’s governance constitution, the system operates entirely without background database servers (e.g., standalone PostgreSQL or Qdrant containers). This conserves critical GPU and CPU resources, reserving maximum computational bandwidth for LLM inference. All data storage is embedded and in-process.

Storage Matrix

Data Type	Technology	Mode	Details
Relational / State	SQLite	WAL (Write-Ahead Logging)	Stores workspace metadata, agent sessions, task trees, chat messages, and server statuses. Concurrent reads/writes are handled via WAL mode.
Vector RAG	Vector DB	In-process via Apache Arrow	Indexes semantic engineering documents, material specifications, standard parts catalogs, and past design patterns. Runs in-process with minimal memory overhead.
Physical Artifacts	Local Filesystem Vault	Structured Files	Saves binary deliverables (STEP, STL, G-code, FEA meshes) generated by tools. File paths are indexed and queried via SQLite.

Embedded Engine Configurations

SQLite (WAL Mode)

SQLite is configured with PRAGMA journal_mode=WAL; and PRAGMA synchronous=NORMAL; to ensure write operations do not block read processes during concurrent agent runs. The state database contains schemas for: * workspaces: Directories mapped to git repositories. * sessions: Active agent chat history, execution stages, and metadata. * tasks: Hierarchical task breakdowns. * mcp_servers: Installed execution tools and statuses.

Vector DB (In-Process)

The vector store operates in-process utilizing Apache Arrow data structures. It embeds and searches: * Standard CAD libraries (ISO/ANSI parts). * Material properties datasheets. * Internal programming guidelines.

Local Filesystem Vault

A structured directory path vault/ stores binary outputs. The backend maps file paths to specific workspaces, preventing cross-workspace leakage.