RAG & Knowledge· Live independent SaaS build

Orepli

Grounded AI for email and embedded customer chat

A live, multi-tenant customer-support platform that uses one knowledge spine across two channels: an email agent and an embeddable site chat widget. It answers automatically only when evidence clears the configured threshold, otherwise it escalates with a clean summary.

Role · Product and AI architecture

System anatomy

01Grounded RAG

02Confidence Gating

03Escalation

04Multi-Tenant AI

Captured from the live product · October 2026

Open live product

The problem

Customer-support AI needs to be useful without inventing answers, leaking tenant data, or burning through unbounded tokens. Orepli makes grounding, usage limits, escalation, and audit part of the core product model.

What the system does

Capability with an operating model.

Gmail OAuth plus IMAP/SMTP with encrypted credentials

Per-mailbox classify → retrieve → draft → send or escalate pipeline

Embeddable chat widget with origin allowlist and one-time API keys

Three-layer strict-KB refusal contract and multi-turn memory

Knowledge-gap, refusal-rate, and feedback insights for administrators

Shared tenant token pool, hard caps, rate limits, DLQ, and audit timelines

Architecture

How it fits together.

FastAPI + SQLAlchemy + PostgreSQL/pgvector backend

Next.js administration, live preview, and iframe chat experience

Swappable local or hosted embeddings behind a stable retrieval interface

Docker/Nginx deployment with structured logs and exponential LLM retry

Key decisions

Refusal is product behavior

No knowledge, weak retrieval, or an ungrounded final draft all resolve to a safe refusal or escalation.

One usage spine

Email and chat share tenant billing, caps, observability, and administrative controls.

Operate the loop

Feedback and refused questions become visible knowledge gaps rather than disappearing into logs.

Explore another system