SnobBots case study

SnobBots · Ecommerce / Customer support

A multi-tenant AI chatbot platform. Retrieval is scoped to one shop’s content, and new shops are provisioned in bulk.

Private deployment, client project

Each answer is retrieved from one shop’s content, and nothing else.

Embeddings are partitioned per tenant in Pinecone, so retrieval can only reach the requesting shop’s namespace. The model works from one knowledge base, which keeps answers grounded in that shop’s content. Provisioning is automated, so resellers set up shops in bulk instead of one at a time.

What keeps it lean

01Retrieval scoped to one namespace
02Answers grounded in the shop’s own content
03Bulk provisioning instead of manual setup

Each request, user or agent gets the data it needs and nothing more.

The challenge

Resellers wanted to offer AI support agents to their own shops, but every deployment was bespoke, so onboarding one shop took days of manual setup. Worse, a shared index risked one shop’s content surfacing in another shop’s answers. The platform needed to provision shops in bulk and keep each one’s knowledge strictly its own.

The work

We built a multi-tenant MERN platform where each shop runs its own context-aware agent. Embeddings are partitioned per tenant in Pinecone, so retrieval can only reach the requesting shop’s namespace, and answers are grounded in that shop’s catalogue, orders and policies. Automated provisioning sits on top, so resellers set up shops in bulk from a management dashboard.

Inside the product

More than one screen.

3 more screens from SnobBots, captured at device size.

Handover inbox

Conversations the bot could not close, with the order and customer beside them.

Knowledge base

The catalogue, orders API and policies each bot is allowed to answer from.

Shop widget

The customer-facing chat as it appears on a client’s storefront.

On the phone

Home

Handover inbox

Knowledge base

Shop widget

How it is built

Architecture

MERN: React, an Express and Node.js API, and MongoDB for tenant and conversation data. Pinecone holds per-tenant embeddings so retrieval is scoped to one shop’s knowledge base, and provisioning runs without manual setup.

Key features

Multi-tenant architecture with isolated agents

RAG for context-aware, business-specific answers

Automated bulk bot provisioning

Reseller management dashboard

Embeddable support widget

Challenges & solutions

Challenge

Retrieval across a shared vector index surfaced one shop’s content in another shop’s answers.

Solution

Partitioned embeddings per tenant in Pinecone so retrieval can only ever reach the requesting shop’s namespace.

Challenge

Manual per-shop setup made reseller onboarding the bottleneck on growth.

Solution

Automated provisioning end to end, turning a multi-day manual process into a bulk operation.

The impact

What became easier, better or more effective.

Resellers now stand up support agents for their shops in bulk instead of one bespoke deployment at a time. Each bot answers only from its own shop’s content and hands over to a person when it cannot close a conversation, so shoppers get grounded answers and shop owners keep the final word.

60%

faster shop onboarding

Automated provisioning turned a multi-day manual setup into a bulk operation a reseller runs from the dashboard.

isolated namespace per shop

Embeddings are partitioned per tenant in Pinecone, so no bot can ever surface another shop’s content.

100%

of answers grounded in the shop’s own content

Retrieval runs over the catalogue, orders API and policies the reseller has allowed for that bot, and nothing outside them.

Stack