Private Knowledge System
A self-hosted, tier-guarded knowledge store any AI can query — but only through one doorway.
Problem
Personal knowledge scatters across apps and chatbot memories owned by vendors; cloud models can't be trusted with everything, and "trust the client to behave" is not a privacy model.
Solution
One PostgreSQL database on owned hardware holding documents, chunk embeddings, and bitemporal facts — served through a single authenticated FastAPI + MCP doorway that enforces privacy tiers server-side, per client.
Architecture
The knowledge system embodies the portfolio’s central conviction: own the store, rent the machinery. PostgreSQL is the system of record; parsers, embedding models, and even the AI clients are replaceable components around it. Originals are never mutated, facts keep their history instead of overwriting it, and the audit log is deliberately values-free — it records that something happened to a fact, never the fact itself.
Repo evidence
- Six-table schema live — documents, chunks, bitemporal facts, provenance, governance events, review queue
- Hybrid search deployed — BGE-M3 embeddings + cross-encoder reranking, tier-filtered
- Content above a client's tier never leaves the database — enforcement at the doorway, not the model
- Daily pg_dump backups with integrity checks, 14-day retention, restore drill passed