People experiencing homelessness (PEH) face substantial barriers to accessing timely, accurate information about community services, including shelters, food programs, healthcare, and social services, and these gaps are compounded by unstable housing, limited digital access, and the transient nature of available resources. General-purpose LLMs, while conversationally fluent, are prone to hallucination and can confidently return incorrect information about service locations, hours, and eligibility, a failure mode with serious consequences for a population with little margin for wasted trips or wrong information. Reliable, explainable, and up-to-date service navigation is therefore critical infrastructure for supporting PEH and the outreach workers and caseworkers who assist them.
DreamKG addresses this through an agentic, knowledge graph-augmented conversational system that grounds every response in a verified Neo4j knowledge graph of community services across Philadelphia and Los Angeles. The system's agentic design spans three layers: a tool-using router that decomposes multi-service queries, resolves spatial and temporal constraints, and composes multi-stop itineraries; an autonomous ingestion agent that extracts and validates service records from heterogeneous sources; and a bounded entity-resolution sub-agent that investigates ambiguous organization matches, deferring to human review only when evidence is genuinely inconclusive. This keeps the data pipeline auditable and human-gated while reserving autonomous reasoning for query interpretation and ambiguous entity matching, not unsupervised database writes.
We expect this architecture to preserve the reliability advantages shown in earlier evaluation of DreamKG's non-agentic predecessor while improving handling of compositional, multi-step, and ambiguous real-world queries. We anticipate the agentic router will better support multi-turn conversational memory and multi-category requests, and that the entity-resolution sub-agent will reduce duplicate or fragmented organization records relative to threshold-only matching. Evaluation against both the predecessor and general-purpose AI search is ongoing and will be reported as this work continues.