DreamKG: A KG-Augmented Conversational System for People Experiencing Homelessness

Abstract: People experiencing homelessness (PEH) face substantial barriers to accessing timely, accurate information about community services. DreamKG addresses this through a knowledge graph-augmented conversational system that grounds responses in verified, up-to-date data about Philadelphia organizations, services, locations, and hours. Unlike standard LLMs prone to hallucinations, DreamKG combines Neo4j knowledge graphs with structured query understanding to handle location-aware and time-sensitive queries reliably. The system performs spatial reasoning for distance-based recommendations and temporal filtering for operating hours. Preliminary evaluation shows 59% superiority over Google Search AI on relevant queries and 84% rejection of irrelevant queries. This demonstration highlights the potential of hybrid architectures that combines LLM flexibility with knowledge graph reliability to improve service accessibility for vulnerable populations effectively.

• The paper introduces DreamKG, a hybrid system that couples LLM-based query parsing with a Neo4j knowledge graph for service navigation among people experiencing homelessness.
• It employs compositional multi-turn query processing, geocoding, and spatial-temporal filtering to deliver precise, actionable service recommendations.
• Empirical evaluations reveal a 59% improvement in location and temporal query accuracy over generic systems, underscoring its practical utility.

DreamKG: A Knowledge Graph-Augmented Conversational System for Service Navigation Among People Experiencing Homelessness

Problem Motivation and Context

Individuals experiencing homelessness (PEH) encounter considerable barriers to timely, accurate information about essential services such as shelters, food banks, mental health care, and social support, exacerbated by transient living situations, fragmented digital access, and the rapidly shifting landscape of service availability. Conventional LLMs, including ChatGPT and comparable general-purpose systems, while conversationally competent, are characterized by high hallucination rates and unsatisfactory factual reliability, particularly in clinical and resource navigation contexts. Retrieval Augmented Generation (RAG) and KG-based approaches partially address these challenges, but most implementations fail to integrate robust spatial-temporal reasoning or to tailor their architectures specifically for the requirements of community-based service navigation.

System Architecture and Methodology

DreamKG introduces a hybrid architecture coupling LLM-based natural language query understanding with a Neo4j-implemented knowledge graph that encodes structured, community-validated service information for the Philadelphia metropolitan area. The system pipeline includes: (1) detailed keyword extraction with support for service types, location descriptors, temporal constraints, and conversational cues; (2) input normalization and geocoding for robust spatial reasoning; (3) Cypher-based schema-constrained query generation, employing dynamic fallback mechanisms for ambiguous queries; (4) KG retrieval with downstream spatial (distance) and temporal (operating hours, day-of-week) filtering; (5) hierarchical formatting of output, integrating both succinct actionable directions and rich organizational context; and (6) multi-modal response rendering, including embedded maps and service cards.

A key architectural innovation is compositional query construction enabling decomposition and resolution of complex, multi-turn, context-aware user queries. The KG schema is ontology-driven, supporting five essential service domains (mental health, social security, libraries, shelters, food), and integrates heterogeneous data from authoritative public and governmental datasets.

Key Features and User Interaction

DreamKG’s user interface, implemented in Streamlit, supports diverse query modalities, explicit and implicit spatiotemporal expressions (e.g., “near me”, “after 8pm”), and multi-turn dialogue memory enabling reference resolution and context retention across conversational states. Location-aware ranking, operational hour filtering, interactive geospatial visualizations, and structured details for all service entities ensure high actionability and transparency. The system logs all queries and decision traces to facilitate ongoing auditability and improvement.

Empirical Evaluation and Results

DreamKG was benchmarked through a structured, blind, pairwise evaluation using 300 queries (200 domain-relevant, 100 irrelevant), judged by GPT-4.1-mini according to explicit criteria: location specificity, domain relevance, operational detail, structural clarity, actionability, signal-to-noise ratio, and overall accuracy. Across relevant queries, DreamKG exhibited a 59% superiority over Google AI Mode, and for irrelevant queries, demonstrated an 84% rejection rate versus potentially spurious or inapplicable results. Strong performance was particularly evident in location-sensitive and temporally constrained queries, where DreamKG’s spatial-temporal filtering produced more precise, actionable recommendations than LLM-based or generic search alternatives.

A documented limitation is DreamKG’s dependence on scheduled data for real-time operational status, constraining adaptability to unscheduled closures or dynamic disruptions, where baseline systems occasionally performed better due to their broader data aggregation.

Novel Contributions

DreamKG advances KG-augmented conversational AI through three core innovations:

1. Compositional multi-turn query processing enabling robust conversational reference tracking and dynamic decomposition of complex search intents.
2. Constrained, ontology-driven KG retrieval architecture integrating geocoding, coordinate-based pruning, and fine-grained temporal windowing across heterogeneous service databases, with rigorous error handling for user and data ambiguities.
3. Community-curated, modular database construction, facilitating ongoing extension and responsiveness via active integration with local stakeholders and resource providers.

These features collectively yield a framework optimized for question understanding and real-world service navigation challenges faced by PEH.

Theoretical and Practical Implications

DreamKG substantiates the empirical superiority of KG-augmented conversational architectures over purely LLM-based or unstructured search paradigms for structured, high-stakes information retrieval in community care domains. The results reinforce the necessity of explicit spatial-temporal modeling and deep integration of external knowledge representations for high-consequence deployments. The system’s transparent logging and modularity recommend its adoption in other urban and social care contexts, provided that local data collection and ontology adaptation are feasible.

Future research trajectories include expansion to multi-site deployments, incorporation of real-time event feeds (e.g., unexpected closures, program changes), deeper personalization features, and closed-loop evaluation with target populations to ascertain impact on resource access, case management, and downstream social/health outcomes.

Conclusion

DreamKG operationalizes a robust KG-augmented conversational system specifically engineered for people experiencing homelessness, demonstrating significant performance advantages over generic LLM-based alternatives in spatial-temporal service navigation. The hybrid architecture, compositional query modeling, and community-driven curation pipeline offer a transferable blueprint for high-reliability, actionable AI in service provision for vulnerable populations. Continued evaluation and refinement across diverse settings will be necessary to maximize real-world impact and support equitable access to essential community resources.

Reference: "DreamKG: A KG-Augmented Conversational System for People Experiencing Homelessness" (2604.11703)