QSR Assistant

Institutional memory

Every admin correction becomes a permanent, versioned knowledge artifact. The system grows smarter from its own mistakes — a closed feedback loop that most enterprise RAG deployments skip entirely.

→ self-improving system

Autonomous document ingestion

Upload a PDF. The system chunks it, parses it, embeds it, and runs a 30-question synthetic eval to verify retrieval quality — all without human intervention. Non-technical operators can safely add new documents.

→ self docs ingestion

Document-agnostic vertical framework

The entire stack — RAG engine, admin, Gold Store, evaluation suite — is structurally independent of the CCNL. Swap the corpus and the same product serves procedures and manuals vertically.

→ vertical AI framework

Data sovereignty

Today: only context fragments reach Anthropic under the zero-retention agreement. Future development could run Llama 3 locally, closing the loop entirely and converting to a 100% local approach.

→ data sovereignty

Telegram human-in-the-loop escalation

When confidence drops below threshold, the system escalates to the admin via Telegram in real time. No manager is ever left with a wrong answer that nobody knows about.

→ human oversight layer

Evaluation suite as deployment gate

A 30-question synthetic evaluation (faithfulness, relevance, context precision, recall) runs after every document ingestion. You can't silently break the system by uploading a bad PDF.

→ rag_evaluator