Knowledge Base
LLM-built wiki from transcripts, emails, and chat
What it does
The knowledge base accepts structured items from the input pipeline (notes, todos, CRM facts) and stores them by topic and language. It is the storage layer behind "what does Reinier already know about X?"
Two shapes coexist
- Fleeting notes (today) — append-only atoms grouped by source, captured exactly as the pipeline emits them. Cheap, high-recall.
- Evergreen wiki pages (future) — per-entity living documents synthesised from the fleeting notes. Higher cost, higher signal — these are what the Papegaai use-cases page reads from at build time.
Access layers
The access model is three stages × principals: raw artefacts, the inbox (pre-routing, source-language, source-grouped), and notes/wiki pages (per-KB language, post-promotion). Humans get a web UI; agents get bearer-token APIs.
Web routing
kb.papegaai.ai → Cloudflare DNS (grey) → Caddy on the input-pipeline VPS → the KB interface service. Authentication uses the same Google-OAuth + email-allowlist pattern as the rest of the Papegaai fleet.
Status
In dev. Inbox + notes are live. Wiki-page synthesis is the next major surface.