Facilities Voice Agent with RAG Troubleshooting

The Challenge

When a piece of equipment fails at a restaurant location — a fryer stops heating, a freezer starts temping, a glazer malfunctions — the store manager needs to act fast. But the process for getting help was friction-heavy: log into the facilities management platform, find the right equipment, categorize the issue, create a work order, and hope it gets routed to the right vendor.

Many managers would just call someone instead. The problem was that “someone” was a small facilities team that couldn’t scale to handle calls from hundreds of locations, especially during peak failure times.

The Approach

I built a voice agent that store staff can call to report equipment issues, get walked through troubleshooting steps, and automatically create work orders when on-site resolution isn’t possible.

The voice agent handles the full call flow:

1. Caller identification — looks up the caller against the employee directory to determine their role and location
2. Equipment identification — the caller can scan an NFC tag on the equipment or describe what they’re looking at; the agent matches it against the location’s asset inventory
3. AI troubleshooting — before creating a work order, the agent consults a knowledge base of repair procedures and walks the caller through relevant steps
4. Work order creation — if troubleshooting doesn’t resolve the issue, the agent creates a properly categorized work order in the facilities management platform with all the context from the call
5. Post-call analytics — every call generates structured data (store, equipment, troubleshooting attempted, outcome) that feeds into operational dashboards

The troubleshooting knowledge base is a RAG system built on repair videos and service manuals from multiple restaurant brands. The ingestion pipeline transcribes repair videos, extracts structured procedures using AI, and stores them as vector embeddings in a pgvector database. When a caller describes a symptom, the system finds the most relevant repair procedures using semantic similarity search, with confidence scoring to decide whether to present a procedure directly or ask the caller to confirm.

The architecture is three MCP servers working together:

• Facilities MCP — 23 tools covering the full work order lifecycle (create, assign, escalate, complete), equipment queries, and location lookups. Built on a custom Python SDK I wrote for the facilities management API.
• Troubleshooting MCP — vector search over repair procedures, with brand and equipment type filtering
• Employee/data MCP — caller lookup, store information, and configuration data

All three are independently deployed on Vercel and orchestrated by the voice agent’s LLM reasoning. The voice infrastructure handles transcription, text-to-speech, and phone connectivity, while the agent focuses purely on the conversation logic and tool usage.

The Results

Store staff can now call a phone number, describe their equipment problem in natural language, and either get walked through a fix or have a properly documented work order created — all without touching the facilities management platform directly.

The troubleshooting step is the highest-value part of the system. When a repair procedure exists in the knowledge base and the caller can follow the steps, it eliminates the need for a vendor dispatch entirely. Even when a work order is needed, the call transcript provides the vendor with detailed context about what was already tried, reducing diagnostic time on-site.

The system processes calls across multiple restaurant brands, each with its own equipment types and repair procedures, using the same underlying architecture with brand-specific knowledge bases.