Skip to content

Offline Caching Strategy

To facilitate complete offline operation in air-gapped or restricted-network environments, Wright implements a write-through offline caching system for both external Model Context Protocol queries and LLM contexts.


1. Local Package Caching

When running behind a firewall or offline: * Apt Packages: Pre-cached in wright_varcache volume during the initial container build. * Python Dependencies: Local packages can be installed using pre-downloaded wheels stored in a local directory cache or mirrored packages on the host.


2. Vector DB Metadata Caching

For RAG operations: * Engineering manuals, standards PDFs, and dataset vectors are pre-indexed and embedded into the local Vector DB. * Agents search the local in-process DB instead of executing remote cloud embedding lookups.


3. Tool Caching Actions

For third-party integrations (e.g., external REST MCP tools): * A write-through caching layer serializes tool request inputs and responses into a local SQLite table (mcp_cache). * Subsequent requests with identical arguments retrieve the response directly from the local cache, preventing remote network errors.