From Spreadsheet to AI Assistant: How We Built a Self-Updating RAG Agent for Property Queries

I've just deployed Cole Medin's Agentic RAG system for our real estate database, with some custom modifications worth sharing:

What it does: 🔍

📊 Creates a dynamic knowledge base that auto-updates whenever source spreadsheets change

💬 Handles complex natural language queries across your data ("Find products with specific features" or "Compare items matching certain criteria")

📱 Powers customer-facing interactions through QR codes, websites, or messaging platforms

🔗 Returns precisely formatted responses with accurate links and only the most relevant details

📈 Works for any industry with structured data - real estate, e-commerce, inventory management, service catalogs, product specifications, or company knowledge bases

What makes this system awesome: ✨

🧠 It's super smart: It knows when to search for information, when to do math, and when to grab the whole document—just like knowing when to use different tools for different jobs

📚 It reads entire documents when needed: Instead of just looking at random pages, it can read the whole document to understand everything

🔢 It's great at math problems: It can calculate numbers from spreadsheets without making mistakes

🧩 It connects the dots: It can find information that's spread across different files and put it together—like solving a puzzle

📂 It handles lots of files at once: You don't have to upload one file at a time; it works with all your information together

💾 It saves space: It stores all your information in a clever way that doesn't waste computer memory

Technical implementation: ⚙️

🛠️ Built on Assistable using Cole's n8n workflow template

📝 Modified with custom SQL queries for property-specific lookups

🤖 Using ADA-002 for embeddings and GPT-4.0 for responses (Claude Opus seemed slower)

🗃️ Supabase backend with 4000 token chunks and zero overlap

🗣️ Text and voice response formats with separate prompts for each

Challenges faced: 🚧

🔗 Ensuring links remain functional through vectorization

📊 Finding the right chunking strategy for property data

⏱️ Balancing response times (10-30 sec wait currently) with accuracy

Next steps: 🚀

🕸️ Implementing web scraping via sitemap integration

💭 Further refining the conversation history management

If anyone wants to implement something similar for other use cases, I'm happy to help with specific adjustments. Cole's original video provides excellent foundation: https://www.youtube.com/watch?v=mQt1hOjBH9o

Or download his template here: https://github.com/coleam00/ottomator-agents/tree/main/n8n-agentic-rag-agent

#AgenticRAG #AssistableAI #liveknowledge

7 comments