AI-Powered Building Code Lookup

Overview

An n8n workflow that takes a US location and a code type, then returns the currently adopted building code, the adopting authority, the state administrative-code reference, any local amendments, and 3–6 official sources.

The workflow uses an OpenAI function-calling agent (GPT-4o) with the Serper Google Search API as a tool. A separate GPT-4o-mini call resolves ambiguous location strings ("Renton WA", a ZIP code, a full street address) into a canonical "City, State" jurisdiction before the agent runs.

---

How It Works

1. User Input Form — n8n form trigger collects Location and Code Type.
2. Validate Input — code node enforces length (3–200 chars) and allowed code-type values.
3. Resolve City & State — GPT-4o-mini (temperature 0) resolves the raw location into City, Full State Name, or returns UNKNOWN.
4. Parse City/State — falls back to the raw location if resolution returns UNKNOWN.
5. Build Agent Prompt — composes the research brief for the agent.
6. Building Code Research Agent — GPT-4o agent that calls the web_search tool (Serper) up to 5 times following a deliberate search strategy (city ordinance → state administrative code → local amendments).
7. Structured Output Parser — enforces the JSON schema for downstream rendering.
8. Format HTML Output — code node escapes user content and renders a safe HTML summary.
9. Show Results to User — n8n form completion page displays the formatted summary.

---

Form Inputs

Field	Type	Notes
Location	text (required)	City + State, ZIP, or full address. Example: Renton WA, 98168, 4414 S 144th St.
Code Type	dropdown (required)	One of: Building Code, Residential Code, Fire Code, Mechanical Code, Plumbing Code, Electrical Code.

---

Output Schema

The agent returns JSON conforming to:

{
  "executive_summary": "2–3 sentence overview of the current code status.",
  "base_code": "e.g., 2021 Uniform Plumbing Code",
  "adopting_authority": "e.g., Washington State Building Code Council",
  "state_code_reference": "e.g., Chapter 51-56 WAC",
  "local_amendments": "Summary of city-specific amendments, or 'None found'.",
  "results": [
    {
      "title": "...",
      "link": "https://...",
      "snippet": "...",
      "source_type": "city | state | publisher | other"
    }
  ]
}

The results array contains 3–6 official sources, prioritized in this order: official municipal / .gov / .us domains → recognized publishers (codepublishing.com, municode.com, ecode360.com, iccsafe.org) → state agency PDFs.

---

Setup

1. Credentials

Two n8n credentials are required.

OpenAI — already wired up in the workflow under the name OpenAi account. Open the Resolve City & State and OpenAI Chat Model (GPT-4o) nodes and confirm the credential is selected.

Serper API — create a new credential in n8n:

• Type: HTTP Header Auth
• Name: Serper API
• Header name: X-API-KEY
• Header value: your Serper API key (get one at serper.dev)

Then open the Web Search Tool (Serper) node and select the Serper API credential. The exported workflow has REPLACE_WITH_SERPER_CREDENTIAL_ID as a placeholder until you do this.

2. Import the Workflow

1. Open your n8n instance.
2. Workflows → Import from File → upload AI-Powered Building Code Lookup.json.
3. After import, open the two OpenAI-backed nodes and the Serper node and verify credentials are attached.
4. Activate the workflow.

3. Run

Open the form URL exposed by the User Input Form node (path: building-code-finder), fill in a location and code type, and submit. The result page renders inline.

---

Models & Cost Profile

Step	Model	Why
City/state resolution	gpt-4o-mini (temp 0)	Cheap, deterministic geocoding.
Research agent	gpt-4o (temp 0.1)	Function-calling support and reliable grounding.
Web search tool	Serper (google.serper.dev/search)	Up to 5 calls per run; agent decides when to stop.

---

Architectural Notes

• The agent node (@n8n/n8n-nodes-langchain.agent) is required for tool calling — attaching tools to a plain LLM completion node leaves them uninvoked.
• The Structured Output Parser is wired as the agent's ai_outputParser, so the HTML node always receives the expected shape.
• The HTML formatter escapes all user/AI strings and validates URLs against ^https?:// before rendering, to keep the form completion page safe from injection.
• The agent system prompt forbids fabricated URLs — every link in the output must come from a web_search tool result in that run.

---

Files

• AI-Powered Building Code Lookup.json — the n8n workflow export.
• User_Instructions.md — end-user guide for running the form.
• Outbrief_Session.md — session debrief and next-step recommendations.

---

© 2026 Greyfibre Developer Network | Prepared by Bilal Haider