How SOOLDD Transformed Real Estate Discovery

Rather than a single AI model that attempts to handle all query types, SOOLDD's AI system deploys dedicated agents, each configured to align with a specific interaction context. By coordinating the system with a LangGraph-based orchestration layer, it routes every user query to the right agent. This architecture reflects how SOOLDD's query types vary: a query about a three-bedroom apartment in Austin demands a fundamentally different retrieval and reasoning process than a query about upgrading the subscription plan.

The four specialized agent roles include:

• i Property Inquiry Agent: Handles listing searches, pricing queries, and availability checks with RAG pipeline and SQL agent;
• ii Agent-Matching Agent: Matches renters and buyers with real estate agents based on property type, location, and budget using location-specific intent detection;
• iii Subscription Support Agent: Responds accurately to subscription tier, premium feature queries, and free trial with SOOLDD's knowledge-grounded documentation;
• iv App Navigation Assistant: Guides users through platform features, onboarding flows, and in-app workflows.

Intent routing operates at the query classification layer: the orchestration system organizes every incoming user query by intent rather than by keyword and dispatches it to the appropriate agent. A user who asks, 'How do I upgrade my account?' gets routed to the Subscription Support Agent. A user asking 'show me 3-bed homes in Austin under $3,000 per month' gets routed to the Property Inquiry Agent through the SQL pipeline. Classification is semantic, rather than syntactic.

A multi-agent architecture outperforms single-prompt chatbots in this use case because each agent operates within a specific, scoped context. A unified chatbot that attempts to answer queries related to property, agent matching, navigation, and subscription within a shared context window yields lower accuracy, higher hallucination rates, and inconsistent response quality. Specialized agents eliminate context pollution and enable domain-specific precision in managing each query category.

Before generating each response, the AI system retrieves verified and relevant information from SOOLDD's data, including property listings, agent profiles, and subscription documentation, leveraging the retrieved content to ground the generated response. The large language model generates text based on verified facts rather than invented ones. With this retrieval-first principle, you get an architectural mechanism that minimizes hallucinations, bringing them below 2% in production.

SOOLDD's RAG pipeline uses pgvector, a PostgreSQL extension, deployed alongside the existing MySQL property database. Property listings, agent profiles, and subscription documentation were chunked, embedded using an optimized embedding model calibrated for SOOLDD's real estate domain, and indexed to ensure high-precision retrieval.

The hybrid retrieval approach combines two complementary strategies: vector search (semantic similarity matching) detects conceptually relevant content regardless of how the query is phrased, handling natural language variation that keyword systems can't process. Keyword search (metadata filtering) enables precise filtering based on property type, price range, location, and agent attributes. Both retrieval modes are essential to maintain real estate domain accuracy; vector search manages intent, and metadata filtering manages precision.

Content was classified and chunked, especially to minimize retrieval noise and improve precision for property-specific queries. The chunking strategy was defined based on content type: property listing field groups, agent profiles by attribute clusters, and subscription documentation by feature and pricing tier. This granular chunking proved directly beneficial, achieving a 94.3% pre-launch accuracy benchmark and a sub-2% production hallucination rate.

SOOLDD's MySQL database stores the authoritative structured data for every property listing, with fields including property type, location, price, bedroom count, bathroom count, parking availability, and listing status. The SQL agent converts natural-language search queries into direct, precise, schema-aware SQL queries executed against this live database. A user query like 'Show me 2-bedroom apartments in Brooklyn under $2,500 per month with parking' is parsed by the SQL agent and mapped to SOOLDD's database schema, resulting in a validated SQL query that returns live, verified listing data. The agent works with a grounded understanding of the SOOLDD's table structure, field names, and data types, which prevents schema-mismatch errors that can produce empty/incorrect results.

A validation layer intercepts unsafe, out-of-scope, and ambiguous queries before they reach the implementation, preventing data exposure and erroneous responses. All SQL Agent outputs are validated against a JSON schema before being returned to the NestJS backend, ensuring predictable, structured results that integrate seamlessly with the platform API and render correctly in the mobile app and web interface.

The business impact of this component is significant: it replaced the need for users to navigate complex multi-step filter menus. Natural language took over manual filtering for property searches, reducing the average search time from many minutes of manual navigation to seconds of conversational interaction.

Amazon Bedrock was chosen as the LLM infrastructure layer because it provides managed model hosting, enterprise-level security and compliance, and access to a range of foundational models within a unified AWS framework, without the overhead of model hosting, GPU provisioning, or security management. Considering a growth-stage platform such as SOOLDD, Bedrock offers full-fledged enterprise AI capabilities at an infrastructure cost and operational complexity level that were initially exclusive to large technology organizations.

The AI chatbot layer leverages a secure REST API to integrate SOOLDD's existing NestJS backend. User queries from the JavaScript web platform and the React Native Mobile app are routed to the LangGraph orchestration layer, processed via the relevant agent pipeline, and responses are returned in real time, without needing re-architecture of any existing platform component. Rather than a replacement, the AI layer was engineered to fit as an additional layer of integration.

Production performance is monitored with two observability systems: LangSmith offers AI trace observability, and every agent interaction is logged, scored, and reviewable, ensuring continuous improvement of response accuracy and retrieval quality after deployment. AWS CloudWatch tracks infrastructure performance: availability, latency, and compute metrics. Together, these systems offer the data needed for ongoing AI optimization. The production hallucination rate has remained below 2%, as validated by LangSmith trace analysis.