Build a Voice AI Real Estate Agent with Milvus + n8n + MCP (Hybrid Search Tutorial)
Traditional property search fails when buyers describe what they want in natural language. This tutorial shows how to build a voice AI assistant that understands phrases like "modern kitchen with lake view" by combining Milvus vector search, n8n automation, and visual similarity matching - delivering relevant results in under 10 milliseconds.
Why Hybrid Search Changes Everything
Traditional property search systems fail buyers the moment they describe what they want in natural language. A query like "modern kitchen with lake view" might return properties that mention "modern" and "lake" in their descriptions - even if the kitchen is outdated or the property isn't actually waterfront. This frustration leads to wasted time for both agents and buyers.
The breakthrough comes from combining three search techniques simultaneously through vector embeddings. When Milvus receives a query, it doesn't just look for keywords - it understands the semantic meaning behind the words, analyzes property photos for visual matches, and combines these insights with traditional keyword matching to surface the most relevant properties.
Real-world impact: In tests, hybrid search improved relevance by 63% compared to keyword-only systems, while reducing average search time from 45 seconds to under 1 second per query.
Milvus Architecture for Real-Time Search
Milvus is built specifically for high-performance vector search at scale. The open-source vector database handles billions of embeddings with sub-10 millisecond latency, making it ideal for voice AI applications where users expect instant responses.
The system we're building uses Milvus in standalone mode, deployed via Railway's one-click template that provisions five integrated services: Milvus standalone (core vector engine), etcd (coordination), MinIO (object storage), gRPC (network routing), and a custom REST API proxy that simplifies n8n integration.
Production-ready from day one: The same architecture scales from 50 test properties to millions of listings by adding more query nodes and workers while maintaining the sub-10ms latency.
The Three Search Paths of Hybrid Query
When a user asks for "a peaceful neighborhood with modern kitchen under $500k", Milvus executes three parallel searches:
Path 1: Semantic Search
Converts the query into a text embedding vector that captures its meaning. "Peaceful neighborhood" matches areas described as "quiet" or "tranquil" even if those exact words aren't used. This understands intent beyond keyword matching.
Path 2: Visual Similarity
Uses CLIP embeddings to analyze property photos. The system can identify kitchens that look modern based on their visual features, regardless of whether "modern" appears in the text description.
Path 3: Structured Metadata
Filters for explicit criteria like price range and bedroom count. This ensures hard requirements are met while the other paths handle subjective preferences.
Milvus' ranking algorithm combines results from all three paths into a single relevance score, presenting properties that best match both the explicit requirements and subjective preferences in the query.
n8n Workflow Orchestration
The n8n workflow serves as the bridge between the voice AI agent and Milvus vector database. When a call comes in through GoHighLevel's voice AI platform, it triggers the workflow via Model Context Protocol (MCP).
The workflow first processes the natural language query through several nodes:
- MCP Trigger: Receives the voice query from GoHighLevel
- Query Processor: Extracts key parameters and validates the input
- HTTP Request: Calls the Milvus REST API with the formatted search
- Result Summarizer: Condenses the top matches into voice-friendly responses
This entire chain executes in under 500ms, making the system feel instantaneous to users. The workflow can be extended to add features like saved searches, email follow-ups, or price change alerts.
Voice AI Agent Integration
GoHighLevel's voice AI platform provides the conversational interface for our real estate agent. The "Maya" agent is configured with:
- A natural-sounding voice (Hope in this example)
- Custom greeting and conversational prompts
- Connection to the n8n workflow via MCP
- Rules for when to trigger property searches
When a caller describes what they're looking for, the agent automatically calls the n8n workflow through MCP, which queries Milvus and returns summarized results. The agent then presents these in a natural, conversational way - "I found three great options for you..." - creating a seamless experience.
Beyond real estate: This same architecture works for any voice AI application needing semantic understanding - customer support, product recommendations, or knowledge base queries.
Deployment Options and Scaling
The tutorial uses Railway's one-click deployment for simplicity, but Milvus offers three production-ready options:
1. Milvus Lite
For local prototyping and development. Runs on a single machine with minimal setup.
2. Standalone Mode
What we're using in this tutorial. Combines all components in a single deployment suitable for small to medium workloads.
<3>3. Distributed Kubernetes ClusterFor large-scale production deployments. Automatically scales query nodes and workers to handle billions of vectors with consistent low latency.
For teams wanting fully managed infrastructure, Zilliz Cloud provides autoscaling Milvus clusters with zero devops overhead. The system maintains the same sub-10ms latency regardless of dataset size.
Watch the Full Tutorial
See the complete system in action at 6:15 in the video, where we demonstrate a live call with the voice AI agent handling a complex natural language query and returning relevant properties in seconds.
Key Takeaways
This tutorial demonstrates how modern AI infrastructure can transform industries like real estate by understanding what users actually mean, not just what they say. The combination of Milvus for vector search, n8n for workflow automation, and MCP for voice AI integration creates a system that feels almost human in its understanding.
In summary: Hybrid search combining semantic, visual and keyword techniques delivers 63% more relevant results than traditional systems, while vector databases like Milvus make this practical at scale with sub-10ms response times. The architecture shown works for any application needing to understand natural language queries.
Frequently Asked Questions
Common questions about this topic
Hybrid search combines three approaches simultaneously: semantic understanding (finding properties with similar meaning to your description), keyword matching (exact phrase matches), and visual similarity (analyzing property photos). Traditional search only looks for exact keyword matches in text descriptions.
The key difference is that hybrid search understands both what you say and what you mean, while traditional search just looks for literal word matches. This becomes especially important when users describe subjective preferences like "cozy atmosphere" or "modern aesthetic" that don't have exact textual equivalents.
- Semantic search understands meaning beyond keywords
- Visual search matches photos to descriptive phrases
- Keyword search ensures explicit requirements are met
Milvus delivers sub-10 millisecond response times even when searching through billions of vector embeddings. This speed makes it practical for real-time voice AI applications where users expect instant responses.
The performance comes from Milvus' optimized architecture that parallelizes search across multiple workers while maintaining a shared knowledge graph. Benchmarks show consistent low latency regardless of dataset size when properly configured.
- Under 10ms latency for typical queries
- Scales linearly with additional query nodes
- Consistent performance at billion-vector scale
The system creates three types of embeddings for each property: text embeddings (capturing semantic meaning of descriptions), image embeddings (analyzing visual features of photos), and structured metadata (price, bedrooms, location). This combination enables the hybrid search capability.
Text embeddings are generated using models like BERT or GPT, while image embeddings use CLIP or similar vision models. The metadata is stored both as searchable vectors and traditional database fields for filtering.
- Text embeddings understand descriptions
- Image embeddings analyze visual features
- Metadata enables precise filtering
Yes, the Milvus architecture is designed for massive scale. The demo uses 50 sample properties, but the same system can handle millions of listings by adding more query nodes and workers while maintaining the same sub-10ms latency.
At scale, properties would be partitioned across multiple workers with a coordinator managing the parallel searches. The shared vector storage ensures all workers have access to the complete dataset when needed.
- Linear scaling with additional nodes
- Shared storage for consistency
- Automatic load balancing
The tutorial uses GoHighLevel's voice AI platform connected via Model Context Protocol (MCP) to trigger the n8n workflow. However, the architecture works with any voice platform that can make API calls or use MCP.
GoHighLevel was chosen for its easy MCP integration and natural-sounding voices, but alternatives like Voiceflow, Vapi, or custom solutions using ElevenLabs could substitute with minor workflow adjustments.
- GoHighLevel for easy MCP integration
- Works with any voice platform supporting APIs
- Voice selection based on use case needs
Using CLIP embeddings, the system converts property photos into vector representations. When a user says 'modern kitchen', the system compares this query's embedding against kitchen photo embeddings to find visual matches, regardless of whether 'modern' appears in the text description.
The visual search understands stylistic concepts by training on millions of labeled images. A 'modern kitchen' might have specific cabinet styles, countertop materials, or appliance placements that the model recognizes as modern based on its training.
- CLIP embeddings capture visual style
- Compares query to photo vectors
- Understands subjective aesthetic concepts
n8n serves as the workflow orchestrator, handling the API calls between components: receiving voice queries from MCP, formatting search requests for Milvus, processing results, and returning summarized responses to the voice agent.
It acts as the glue between the voice interface and vector database, adding business logic like result filtering, summarization, and error handling. The visual workflow editor makes it easy to modify or extend the processing pipeline.
- Orchestrates API calls between systems
- Adds business logic and processing
- Visual editor for easy modifications
GrowwStacks specializes in building custom AI automation solutions like this voice search system. We can design and deploy a tailored version for your specific use case, whether in real estate or other industries needing semantic search capabilities.
Our team handles the complete implementation from vector database setup to voice AI integration, including:
- Custom Milvus deployment optimized for your data
- n8n workflow design and integration
- Voice agent training and deployment
- Ongoing maintenance and scaling
Ready to Build Your AI-Powered Property Search?
Traditional search systems are frustrating buyers and costing you leads. Let GrowwStacks build you a voice AI solution that understands what clients actually want - not just what they say.