HabitatAgent: An End-to-End Multi-Agent System for Housing Consultation

Battle Count: 67

This paper is focused on housing consultation and real estate decision support. While it uses LLMs, GraphRAG, and multi-agent architectures that could theoretically be adapted to financial domains, the paper itself has no direct relevance to quantitative trading, portfolio optimization, or market prediction. The methodology patterns (verification-gated memory, adaptive retrieval, multi-tier validation) could inspire similar architectures in financial decision support, but the paper does not address any trading-related problems.

The system involves four specialized agents with complex internal workflows: a four-layer memory system with verification-gated updates, adaptive retrieval routing with multiple backends (Milvus vector DB, Neo4j graph DB, MySQL, BM25/jieba), 14 task-specific prompt templates, multi-tier validation (factual consistency, entity accuracy, compliance), and failure-type-aware remediation. Requires integration of multiple databases, LLM orchestration, knowledge graph construction, and real-time compliance checking. Production deployment adds further complexity with streaming generation, async memory updates, and multimodal output.

The paper uses a proprietary dataset from a Beijing housing platform (Fangdongdong) with 5,000 property listings and 100 real consultation scenarios. No public code or data repository is mentioned. The system architecture is described in detail, but the proprietary nature of the data and production system limits reproducibility. The knowledge graph (6,016 nodes, 45,000 edges) and specific prompt templates are not publicly available.