Sbc

An end-to-end pipeline that lets you query the Saudi Building Code (SBC 201-CC-2024) using natural language and get grounded, section-referenced answers — powered by Agentic Hybrid RAG served via MCP in a multi-agent architecture.

What It Does

Ask a question like "What is the minimum tread depth?" and get:

275 mm (Section 1011.5.2, pages 1106–1107)

No hallucinations. Just the actual code, with section references.

Architecture

Colab (processing + embeddings)
    → Supabase (vector DB + keyword search)
        → n8n (AI Agent #1: Agentic Hybrid RAG)
            → MCP (protocol layer)
                → Claude (AI Agent #2: conversational interface)

Why "Agentic Hybrid RAG with MCP"?

• Hybrid RAG — Two retrieval methods: vector search (semantic similarity) + keyword search (exact term matching). One knowledge base.
• Agentic — The n8n AI Agent autonomously decides how to query both retrieval tools, combines results, and generates grounded answers. It reasons over what it finds — not just retrieve-and-paste.
• MCP — The workflow is exposed as an MCP server, so Claude can connect to it as a second AI Agent. Two agents, two roles, one pipeline.
• Multi-Agent — Claude handles conversation and user interaction via MCP. The n8n AI Agent handles retrieval and code-grounded reasoning.

Stack

Layer	Tool	Role
Data Processing	Google Colab	PDF extraction, chunking, embedding generation
Embeddings	OpenAI (text-embedding-ada-002)	Vector embeddings for semantic search
Knowledge Store	Supabase	Vector database (match_sbc_documents) + keyword search (keyword_search_sbc)
AI Agent / Orchestration	n8n	AI Agent with dual retrieval tools + direct chat interface
LLM	GPT-4.1-mini	Reasoning and answer generation
Protocol	MCP	Exposes the n8n workflow to external AI clients
Interface	Claude.ai	Conversational front-end via MCP connection

Repository Contents

├── README.md
├── SBC_RAG_Pipeline.ipynb          # Colab notebook: PDF processing, chunking, embedding, upload
├── SBC_MCP_Query.json              # n8n workflow: AI Agent with dual retrieval + MCP
└── screenshots/                    # Architecture diagram and demo screenshots (optional)

Setup Guide

1. Supabase

Create a Supabase project and set up the following:

Vector table — sbc_documents:

• id (int8, primary key)
• content (text)
• metadata (jsonb)
• embedding (vector(1536))

Functions:

• match_sbc_documents — vector similarity search using pgvector
• keyword_search_sbc — full-text keyword search on content

2. Colab Notebook

Open SBC_RAG_Pipeline.ipynb in Google Colab:

1. Upload the SBC 201-CC-2024 PDF
2. Set your environment variables:
   • OPENAI_API_KEY
   • SUPABASE_URL
   • SUPABASE_KEY
3. Run all cells — this will:
   • Extract text from the PDF
   • Chunk it into meaningful sections
   • Generate OpenAI embeddings
   • Upload everything to Supabase

3. n8n Workflow

Import SBC_MCP_Query.json into your n8n instance:

1. Go to n8n → Import Workflow → paste the JSON
2. Configure credentials:
   • OpenAI — your API key
   • Supabase — your project URL and API key
3. Update the HTTP Request node with your Supabase URL and key
4. Activate the workflow
5. Enable "Available in MCP" in workflow settings

4. Claude.ai (MCP Connection)

1. In Claude.ai settings, add your n8n MCP server URL
2. Start a new conversation and ask any SBC question

Demo Questions

These queries consistently return strong, grounded answers:

• "What is the minimum tread depth?"
• "What is the maximum riser height?"
• "What is the minimum stair width?"

Key Decisions

• Dual retrieval — Vector search alone misses exact terms; keyword search alone misses context. Using both gives the most reliable results.
• topK=12 — Retrieves 12 chunks per vector query for sufficient coverage of complex code sections.
• GPT-4.1-mini — Fast and cost-effective for the RAG reasoning layer. Claude handles the conversational interface.
• Session memory — The n8n agent maintains conversation context within a session for follow-up questions.

Limitations

• Some SBC sections (e.g., fire rating for exit stairway enclosures) are not reliably retrieved — a known chunking/embedding gap that could be improved with better chunking strategies or contextual enrichment.
• The system works best with specific, measurable code queries (dimensions, heights, widths) rather than broad conceptual questions.

License

This project is for educational and research purposes. The Saudi Building Code (SBC 201-CC-2024) content is not included in this repository due to copyright.

Author

Built as part of the MSc in AI for Architecture, Engineering, Construction & Operations (MAICEN) program at Zigurat Institute of Technology.

---

This is one approach to grounded AI for AEC professionals — building codes that answer back.