Reference: Enterprise-Scale Knowledge Agent

Architectural notes

Reference architecture for a knowledge agent at enterprise scale —
millions of documents across Confluence, SharePoint, code repos, with
per-user access enforcement and audit logging.

Why these picks:
- LangGraph for explicit access-check + retrieval + synthesis flow.
- Pinecone over turbopuffer for mature enterprise features (BYOC, RBAC).
- Self-hosted Langfuse for data sovereignty over query logs.
- Per-user ACL filter applied at retrieval time — never rely on the LLM.

Common gotchas:
- ACL leakage via embedding mixing is the existential risk; separate
  indexes per security boundary.
- Confluence/SharePoint sync lag causes "agent says the wrong thing"
  incidents; surface freshness explicitly.