AI Knowledge Architecture

Architecture determines whether governance is enforceable or performative. A RAG demo wires one index to one bot. Production needs many apps, many collections, shared sources, and hard isolation between audiences. AI knowledge architecture maps those relationships before code hardens the wrong boundaries.

Bad architecture shows up late: you cannot separate customer and internal retrieval without re-indexing everything. You cannot trace citations because chunks lost document lineage. Fix architecture early or pay migration tax later.

Model entities: sources, collections, applications, policies, indexes, and evaluation suites. Draw data flows from CMS and policy systems through ingestion, governance, retrieval, and generation.

Choose separation strategies: workspace isolation, collection-scoped indexes, metadata gates, or hybrid. Document precedence rules where collections overlap.

Plan for observability hooks at retrieval and policy decision points. Architecture should make debugging a wrong answer a data query, not a log archaelogy project.

Nintendo architecture defines three apps sharing one Help Center source but different collection scopes. The chatbot index excludes internal-only metadata fields. The sales copilot adds a second index bound to internal collections with separate embedding namespaces so filters cannot leak across boundaries.