LLM-Based Re-Ranking for Real Estate Search

Research Context and Motivation

Finding a home is a complex, high-stakes process involving multi-dimensional preferences—such as lifestyle needs, neighborhood character, and conditional trade-offs—that are difficult to capture with traditional filter-based search. As users increasingly shift toward conversational interfaces, there is a growing need for systems that can interpret natural language intent rather than relying on rigid, structured constraints. This paper addresses this challenge by integrating an LLM-based re-ranker into the 'Concierge' platform at QuintoAndar, a leading Latin American housing marketplace.

The Proposed Approach

The authors implement a two-stage recommendation pipeline. First, a traditional retrieval system generates a candidate set based on structured filters. Second, an LLM-based re-ranker refines this list by scoring each candidate independently. The re-ranker uses a 'point-wise' scoring strategy, which allows for parallel processing to maintain low latency. The scoring process integrates four key inputs: the user's current natural-language request, a long-term textual user profile (derived from historical interactions), structured filters, and aggregate statistics about the entire candidate set (e.g., price distributions). By including candidate-set statistics, the model can judge whether a property is a good value relative to others, rather than evaluating it in isolation.

Evaluation and Results

To validate the system, the researchers constructed a large-scale dataset of 960,000 query-item pairs, annotated using an 'LLM-as-a-Judge' framework with human validation. Offline testing confirmed that the re-ranker consistently improved ranking quality. More importantly, a production A/B test showed that the model significantly outperformed the baseline, yielding a 5.3% increase in click-through rates and a 4.8% increase in scheduled visits. These results demonstrate that LLMs can effectively bridge the gap between natural language intent and structured property metadata without requiring a complete overhaul of the existing search infrastructure.