We are developing a plug-and-play multitenant AI chatbot platform for furniture e-commerce companies. This chatbot will act as a furniture industry expert plus a personalized company expert for each tenant's e-commerce store. Once a company provides their website link, the platform will scrape and ingest product data, images, prices, and other details, and automatically train a dedicated chatbot instance.
This AI assistant will:
Answer customer questions using scraped data from the website.
Act as a furniture industry expert to help a customer choose products.
Be deployable in various formats: chatbubble, API, or embeddable widgets.
Provide fast and frictionless integration for businesses of any size.
By offering immediate, intelligent support, this tool will reduce bounce rates, guide customers to the right product, and convert more leads into sales. Tenants also gain access to valuable data insights from user queries.
Main Features of the Chatbot
Multimodal Model : The system will read and extract data from tenant websites including images, product specs, tables, and PDFs to fine-tune the chatbot per tenant.
Hybrid Control : Chatbot will escalate conversations to live agents (via webhook/API to tenant CRM or live chat platform) in defined scenarios like complex requests, refund issues, or complaints.
Chatbubble : Easy-to-embed JavaScript snippet for tenants to add chatbot as a chatbubble on their site.
Tenant Dashboard : Tenants will log in to a dashboard to -
Pay As you go model : The tenants can buy message credits/tokens based on the customer volume instead of the need to buy a specific monthly plan. Pricing will adjust dynamically on the basis of number of user messages.
Multitenancy : Separate database schemas or logical isolation per tenant. Each tenant gets an isolated chatbot, independent training, and seperate token usage.
Context History : The chatbot will retain session-based user context for smooth conversations.
Recording Chat history : Logged chats are accessible to tenants via dashboard for customer insights. Can also give the ability of exporting the chat logs via CSV
What all we did
System ideation and architecture.
Model training, multimodal web scraping
Tenant dashboards.
Token-based usage tracking.
Seamless frontend + backend integration.
Deployment, documentation, and support.
Target Tenants:
Furniture e-commerce stores
Furniture brands
Furniture aggregation platforms
Technical Framework
Tenant Creation & Isolation
Each tenant had a unique namespace in the vector database (Pinecone).
Data isolation was maintained across tenants.
Website Scraping
Tenant-provided websites were scraped using Crawl4AI.
Content was cleaned, chunked, and converted into embeddings.
Vector Storage
Embeddings were stored in Pinecone under tenant-specific indexes.
RAG Pipeline
User inputs were embedded using OpenAI's text-embedding-ada-002.
Similar documents were retrieved from Pinecone.
Retrieved context + query were sent to OpenAI GPT (gpt-3.5/4) for generating a response.
.png)
