Production Knowledge System

Building a production-grade knowledge system forces you to confront problems that most RAG tutorials skip entirely: how do you handle knowledge that changes over time? What happens when a new document contradicts an older one? How do you ingest from heterogeneous sources — documents, notes, conversations, code — without writing a custom parser for each? The architecture centers on three problems: ingestion normalization (a common schema regardless of source format), memory consolidation (detecting and resolving when new knowledge supersedes older entries), and query routing (combining semantic similarity with structured retrieval). The ingestion pipeline processes raw content through a normalization layer before embedding, so vector search operates over clean, standardized knowledge units rather than raw document chunks. Content-addressed storage using hashes prevents duplicate re-embedding and enables efficient incremental updates. The architecture is fully designed and documented; the core ingestion pipeline and embedding layer are built. Memory consolidation — the hardest piece — is under active development. The design work came directly from shipping a production knowledge system in a commercial context, and the personal build is the space to solve the remaining problems properly, without production pressure narrowing the solution space. This project makes a specific claim: senior agentic engineering is about knowing which problems are solved and which require real design work. The solved problems — chunking, embedding, vector retrieval — are handled by existing tooling. The unsolved problems — supersession detection, cross-source consolidation, knowledge freshness — are where the actual engineering happens. This build makes that distinction visible.