TrueTerra Oils AI Chatbot

Overview

TrueTerra Oils AI Chatbot is an intelligent, customer-facing wellness assistant named 'Terra' that provides personalized essential oil recommendations and guidance based exclusively on verified TrueTerra product data. Built with a RAG architecture powered by GPT-4o-mini and a 2,446+ document knowledge base, Terra delivers conversational, FDA-compliant responses 24/7 — helping customers discover products, understand usage, and navigate an extensive catalog of essential oils and wellness items.

Key Features

• Conversational AI assistant 'Terra' with personalized essential oil recommendations
• RAG architecture with 2,446+ ingested markdown knowledge base documents
• Product catalog integration from structured Excel data (ProductCatalog.xlsx)
• Semantic vector search via MongoDB Atlas and OpenAI text-embedding-3-small
• Persistent conversation thread management with history
• React/TypeScript frontend with real-time chat UI and markdown rendering
• In-memory caching layer (5-minute TTL) for improved response performance
• FDA-compliant language enforcement across all AI-generated responses
• JWT authentication and CORS security on all API endpoints

Challenges

• Processing and structuring 2,446+ markdown files into a searchable vector knowledge base
• Enforcing FDA-compliant language consistently across all AI responses
• Building a personalized recommendation engine from a large product catalog
• Maintaining conversation context across multiple threads per user
• Delivering fast responses from a large vector database at scale

Solutions

• Built automated ingestion pipeline with text chunking (2000 chars) and OpenAI embedding generation
• Engineered system prompts with strict FDA compliance rules and brand voice guidelines
• Parsed ProductCatalog.xlsx into structured MongoDB documents for semantic product matching
• Implemented persistent conversation storage per user with TanStack Query for state management
• Added NodeCache in-memory caching layer with TTL to reduce repeated vector search latency