The Challenge: Buyers Search for Lifestyles, Portals Offer Filters
A UK property search platform lists 180,000 properties across sales and lettings. Their search interface offers standard filters: location, price range, bedrooms, property type. But user research reveals that buyers think in lifestyle terms: “a period cottage with a large garden within 30 minutes of Bristol, under £400k, in a village with a pub and primary school.” Translating this intent into filter selections requires multiple searches, manual map checking, and compromise. Analytics show that 55% of users who begin a property search abandon it before saving any results, and exit surveys cite “can’t find what I’m looking for” as the primary reason. The platform loses these potential buyers to estate agents who understand nuanced requirements through conversation.
The platform wants to enable natural language property search — let buyers describe what they want in plain English and receive intelligently matched results. User search queries and browsing behaviour represent commercially sensitive data that reveals market demand patterns; sharing this with third-party AI services would expose competitive intelligence.
AI Solution: LLM-Powered Natural Language Search
A self-hosted LLM via vLLM interprets natural language search queries, extracting structured search parameters (location, radius, price range, property features, lifestyle requirements) and translating lifestyle preferences into queryable criteria. “Quiet village” maps to settlements with population under 5,000 and low road noise index; “good schools” maps to Ofsted-rated outstanding or good primary schools within 2 miles; “near Cambridge” maps to a geographic radius with commute time calculation.
The system combines traditional database filtering with semantic matching — an embedding model scores each property description against the lifestyle elements of the query, surfacing properties that match the feel even when exact filter criteria don’t capture the nuance. Running on a dedicated GPU server, the search responds in under 2 seconds for any natural language query.
GPU Requirements
Natural language search involves two GPU-intensive steps: LLM query parsing (extracting structured parameters from free text) and semantic matching (encoding the query and matching against property description embeddings). Peak traffic of 5,000 searches per hour during Sunday evening browsing requires sustained throughput.
| GPU Model | VRAM | Searches per Hour (7B model) | Response Time |
|---|---|---|---|
| NVIDIA RTX 5090 | 24 GB | ~6,000 | ~1.2 seconds |
| NVIDIA RTX 6000 Pro | 48 GB | ~5,000 | ~1.5 seconds |
| NVIDIA RTX 6000 Pro | 48 GB | ~7,000 | ~1.0 seconds |
| NVIDIA RTX 6000 Pro 96 GB | 80 GB | ~10,000 | ~0.7 seconds |
An RTX 5090 or RTX 6000 Pro handles peak traffic with sub-1.5-second responses. Private AI hosting keeps search query data — revealing real-time market demand — within UK-based infrastructure.
Recommended Stack
- vLLM serving Mistral 7B fine-tuned for property query parsing with structured output.
- Sentence Transformers for encoding property descriptions and query lifestyle elements into semantic vectors.
- Qdrant or Elasticsearch with vector search for hybrid structured + semantic retrieval.
- External data APIs (Ofsted, ONS, Transport for London, National Rail) for enriching searches with school ratings, crime stats, and commute times.
- Conversational refinement — the LLM follows up with clarifying questions when queries are ambiguous.
For analysing property photographs to surface visual features (period features, garden size, kitchen quality), add a vision model. Deploy an AI chatbot for conversational property search as a messaging interface. Use document AI to extract listing data from agent brochures.
Cost Analysis
The 38% improvement in saved search conversion directly impacts the platform’s revenue model (which is driven by agent subscriptions based on buyer engagement). Converting 38% more casual browsers into active searchers generates an estimated £420,000 in additional annual subscription revenue from agents who see higher enquiry volumes. The dedicated GPU server cost is a small fraction of this revenue uplift.
Natural language search also generates rich demand data — the platform learns that 340 buyers per month want “period properties near Stroud with home office space” — intelligence that can be packaged and sold to agents as market insight reports.
Getting Started
Compile 10,000 natural language property search queries from user research sessions and support tickets. Map each to the structured parameters it implies. Fine-tune the LLM on this query-to-parameters dataset. Test against 1,000 held-out queries, measuring parameter extraction accuracy and result relevance. Deploy as an alternative search option alongside existing filters, tracking engagement metrics to quantify the uplift.
GigaGPU provides UK-based dedicated GPU servers for property search workloads. Scale GPU capacity for weekend browsing peaks.
GigaGPU offers dedicated GPU servers in UK data centres with full GDPR compliance. Deploy intelligent search on private infrastructure today.
View Dedicated GPU Plans
