pith. sign in

arxiv: 2604.11703 · v1 · submitted 2026-04-13 · 💻 cs.AI

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

Javad M Alizadeh , Genhui Zheng , Chiu C Tan , Yuzhou Chen , Omar Martinez , Philip McCallion , Ying Ding , Chenguang Yang
show 2 more authors
AnneMarie Tomosky Huanmei Wu
This is my paper

Pith reviewed 2026-05-10 15:24 UTC · model grok-4.3

classification 💻 cs.AI
keywords knowledge graphsconversational AIhomelessnessservice accesshybrid systemsspatial reasoningtemporal filtering
0
0 comments X

The pith

DreamKG augments conversational AI with a knowledge graph to deliver verified information on Philadelphia services without hallucinations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces DreamKG as a hybrid system that pairs large language models with a Neo4j knowledge graph filled with verified details on local organizations, services, locations, and hours. It aims to solve the problem of unreliable answers from standard AI chat systems when people experiencing homelessness seek timely help with community resources. The approach uses structured query understanding to support spatial reasoning for nearby options and temporal filtering for operating hours. A sympathetic reader would care because barriers to accurate information can waste critical time or lead to missed support for a vulnerable population.

Core claim

DreamKG shows that grounding LLM responses in a structured knowledge graph of verified Philadelphia service data enables reliable handling of location-aware and time-sensitive queries, combining the flexibility of conversational models with the factual reliability of graph-based retrieval.

What carries the argument

The Neo4j knowledge graph paired with LLM query understanding, which performs spatial reasoning for distance-based recommendations and temporal filtering for service hours.

Load-bearing premise

The knowledge graph holds verified and current data on Philadelphia organizations, services, locations, and hours, while the test queries represent the actual needs of people experiencing homelessness.

What would settle it

A controlled comparison in which real users experiencing homelessness pose typical queries to both DreamKG and a standard search AI and report which system supplies more accurate, timely, or actionable information.

Figures

Figures reproduced from arXiv: 2604.11703 by AnneMarie Tomosky, Chenguang Yang, Chiu C Tan, Genhui Zheng, Huanmei Wu, Javad M Alizadeh, Omar Martinez, Philip McCallion, Ying Ding, Yuzhou Chen.

Figure 1
Figure 1. Figure 1: System Architecture User Query Normalize & Preprocess Extract Keywords Generate Cypher Query Retrieve Information Format Response Output Return Extract Location Location Access Granted? Explicit Location Provided? Geocode Location Use User’s Location Found Within Default Radius? Expand Radius Return Results for Formatting Yes Fallback to Default Location Yes No Relax Search Constraints No No Yes Not Found … view at source ↗
read the original 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 large language models (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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents DreamKG, a hybrid conversational system for people experiencing homelessness that augments LLMs with a Neo4j knowledge graph of Philadelphia organizations, services, locations, and hours. It uses structured query understanding, spatial reasoning for distance-based recommendations, and temporal filtering for operating hours to ground responses and reduce hallucinations. The central claim is that a preliminary evaluation demonstrates 59% superiority over Google Search AI on relevant queries and 84% rejection of irrelevant queries.

Significance. If the evaluation methodology and data verification were robustly documented, the hybrid KG-LLM architecture would offer a concrete example of combining LLM flexibility with structured reliability for high-stakes information access. The explicit support for location-aware and time-sensitive queries is a strength that could generalize to other social-service domains. The work highlights a practical application area but currently lacks the empirical grounding needed for strong impact.

major comments (2)
  1. [Abstract and Evaluation section] Abstract and Evaluation section: the headline claim of 59% superiority over Google Search AI (and 84% irrelevant-query rejection) is presented without any description of the query set size, selection process, representativeness of real PEH needs, exact superiority metric, baseline implementation details, or statistical significance testing. This directly undermines the central empirical support for the system's advantage.
  2. [System Architecture / KG component] System description (KG component): the repeated assertion of 'verified, up-to-date data' on organizations, hours, and locations lacks any account of the verification process, data sources, update frequency, or audit trail. This is load-bearing for the core claim that the KG prevents hallucinations and enables reliable spatial/temporal reasoning.
minor comments (2)
  1. [Abstract and Introduction] The abstract and introduction would benefit from a brief comparison to prior KG-augmented LLM systems (e.g., citations to recent work on retrieval-augmented generation or structured query interfaces).
  2. [Figures and System Overview] Figure captions and system diagrams could more clearly label the flow between LLM query understanding, Neo4j spatial/temporal queries, and response generation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas where additional documentation will strengthen the manuscript. We address each major comment below and will incorporate revisions accordingly.

read point-by-point responses

  1. Referee: [Abstract and Evaluation section] Abstract and Evaluation section: the headline claim of 59% superiority over Google Search AI (and 84% irrelevant-query rejection) is presented without any description of the query set size, selection process, representativeness of real PEH needs, exact superiority metric, baseline implementation details, or statistical significance testing. This directly undermines the central empirical support for the system's advantage.

    Authors: We agree that the current presentation of the preliminary evaluation lacks sufficient methodological detail. In the revised manuscript, we will expand the Evaluation section to describe the query set size, the process used to select queries for representativeness of real PEH needs, the exact superiority metric, the implementation of the Google Search AI baseline, and any statistical significance testing performed. This will provide the necessary transparency and empirical grounding for the reported results. revision: yes

  2. Referee: [System Architecture / KG component] System description (KG component): the repeated assertion of 'verified, up-to-date data' on organizations, hours, and locations lacks any account of the verification process, data sources, update frequency, or audit trail. This is load-bearing for the core claim that the KG prevents hallucinations and enables reliable spatial/temporal reasoning.

    Authors: The referee is correct that the manuscript does not currently detail the verification process for the KG data. We will add a dedicated subsection to the System Architecture description that specifies the data sources, verification steps, update frequency, and audit trail. This revision will directly support the claims about reduced hallucinations and reliable spatial/temporal reasoning. revision: yes

Circularity Check

0 steps flagged

No circularity detected in system description or evaluation

full rationale

The paper is a system-description and preliminary-evaluation manuscript with no mathematical derivations, equations, fitted parameters, or self-citation chains. Claims about spatial/temporal reasoning and the 59%/84% metrics are presented as direct empirical observations rather than quantities derived from the authors' own inputs or prior work by construction. No load-bearing step reduces to self-definition, renaming, or ansatz smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review yields limited visibility into parameters or assumptions; the design implicitly treats the knowledge graph as a reliable external source of truth and assumes LLM query understanding can be made sufficiently structured.

axioms (1)
  • domain assumption A knowledge graph can be kept verified and up-to-date for real-world service data.
    Stated in the abstract as the grounding mechanism but not demonstrated.
invented entities (1)
  • DreamKG hybrid architecture no independent evidence
    purpose: Ground LLM responses in structured service data to reduce hallucinations
    The system itself is the primary contribution; no external falsifiable evidence for its superiority is supplied beyond the preliminary numbers.

pith-pipeline@v0.9.0 · 5460 in / 1329 out tokens · 62331 ms · 2026-05-10T15:24:21.006823+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

  1. [1]

    JMIR Hum Factors 11:e55415

    Radó N, Békási S, Győrffy Z (2024) Health Technology Access and Peer Support Among Digitally Engaged People Experiencing Homelessness: Qualitative Study. JMIR Hum Factors 11:e55415. https://doi.org/10.2196/55415

    work page doi:10.2196/55415 2024

  2. [2]

    JMIR Hum Factors 11:e57114

    Sezgin E, Kocaballi AB, Dolce M, et al (2024) Chatbot for Social Need Screening and Resource Sharing With Vulnerable Families: Iterative Design and Evaluation Study. JMIR Hum Factors 11:e57114. https://doi.org/10.2196/57114

    work page doi:10.2196/57114 2024

  3. [3]

    https://doi.org/10.2105/AJPH.2020.305784

    Benda NC, Veinot TC, Sieck CJ, Ancker JS (2020) Broadband internet access is a social determinant of health! Am J Public Health 110:1123–1125. https://doi.org/10.2105/AJPH.2020.305784

    work page doi:10.2105/ajph.2020.305784 2020

  4. [4]

    J Med Internet Res 24:e38729

    Radó N, Girasek E, Békási S, Gyorffy Z (2022) Digital Technology Access and Health-Related Internet Use Among People Experiencing Homelessness in Hungary: Quantitative Survey. J Med Internet Res 24:e38729. https://doi.org/10.2196/38729

    work page doi:10.2196/38729 2022

  5. [5]

    Omar R, Mangukiya O, Kalnis P, Mansour E (2023) ChatGPT versus Traditional Question Answering for Knowledge Graphs: Current Status and Future Directions Towards Knowledge Graph Chatbots

    work page 2023

  6. [6]

    doi: 10.2196/54345

    Aljamaan F, Temsah M-H, Altamimi I, et al (2024) Reference Hallucination Score for Medical Artificial Intelligence Chatbots: Development and Usability Study. JMIR Med Inform 12:e54345. https://doi.org/10.2196/54345

    work page doi:10.2196/54345 2024

  7. [7]

    A framework to assess clinical safety and hallucination rates of llms for medical text summarisation.npj Digital Medicine, 8:274, 2025

    Asgari E, Montaña-Brown N, Dubois M, et al (2025) A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine 2025 8:1 8:274-. https://doi.org/10.1038/s41746-025-01670-7

    work page doi:10.1038/s41746-025-01670-7 2025

  8. [8]

    JMIR Res Protoc 8:e11022

    Dahrouge S, Gauthier A, Chiocchio F, et al (2019) Access to Resources in the Community Through Navigation: Protocol for a Mixed-Methods Feasibility Study. JMIR Res Protoc 8:e11022. https://doi.org/10.2196/11022

    work page doi:10.2196/11022 2019

  9. [9]

    https://doi.org/101377/hlthaff201901588 39:662–669

    Cartier Y, Fichtenberg C, Gottlieb LM (2020) Implementing Community Resource Referral Technology: Facilitators And Barriers Described By Early Adopters. https://doi.org/101377/hlthaff201901588 39:662–669. https://doi.org/10.1377/HLTHAFF.2019.01588

    work page doi:10.1377/hlthaff.2019.01588 2020

  10. [10]

    npj Health Systems 2025 2:1 2:2-

    Yang R, Ning Y, Keppo E, et al (2025) Retrieval-augmented generation for generative artificial intelligence in health care. npj Health Systems 2025 2:1 2:2-. https://doi.org/10.1038/s44401-024-00004-1

    work page doi:10.1038/s44401-024-00004-1 2025

  11. [11]

    Scientific Reports 2025 15:1 15:40425-

    Wang S, Yang H, Liu W (2025) Research on the construction and application of retrieval enhanced generation (RAG) model based on knowledge graph. Scientific Reports 2025 15:1 15:40425-. https://doi.org/10.1038/s41598-025-21222-z

    work page doi:10.1038/s41598-025-21222-z 2025

  12. [12]

    Scientific Reports 2025 15:1 15:18062-

    Yang Q, Zuo H, Su R, et al (2025) Dual retrieving and ranking medical large language model with retrieval augmented generation. Scientific Reports 2025 15:1 15:18062-. https://doi.org/10.1038/s41598-025-00724-w

    work page doi:10.1038/s41598-025-00724-w 2025

  13. [13]

    J Inf Sci 50:1019–1029

    Rajabi E, Etminani K (2024) Knowledge-graph-based explainable AI: A systematic review. J Inf Sci 50:1019–1029. https://doi.org/10.1177/01655515221112844

    work page doi:10.1177/01655515221112844 2024

  14. [14]

    Fam Community Health 34:72–84

    Vaughn LM, Jacquez F, Zhao J, Lang M (2011) Partnering with students to explore the health needs of an ethnically diverse, low-resource school: An innovative large group assessment approach. Fam Community Health 34:72–84. https://doi.org/10.1097/FCH.0b013e3181fded12

    work page doi:10.1097/fch.0b013e3181fded12 2011

  15. [15]

    J Particip Res Methods 1:2020

    Wallerstein N (2020) Commentary on Community-Based Participatory Research and Community Engaged Research in Health for Journal of Participatory Research Methods. J Particip Res Methods 1:2020. https://doi.org/10.35844/001c.13274

    work page doi:10.35844/001c.13274 2020

  16. [16]

    https://doi.org/101377/hlthaff20151512 35:590–594

    Woolf SH, Zimmerman E, Haley A, Krist AH (2017) Authentic Engagement Of Patients And Communities Can Transform Research, Practice, And Policy. https://doi.org/101377/hlthaff20151512 35:590–594. https://doi.org/10.1377/hlthaff.2015.1512

    work page doi:10.1377/hlthaff.2015.1512 2017

  17. [17]

    https://doi.org/102105/AJPH91121929 91:1929–1938

    MacQueen KM, McLellan E, Metzger DS, et al (2011) What Is Community? An Evidence-Based Definition for Participatory Public Health. https://doi.org/102105/AJPH91121929 91:1929–1938. https://doi.org/10.2105/AJPH.91.12.1929

    work page doi:10.2105/ajph.91.12.1929 2011

  18. [18]

    Soc Work Res 38:59–64

    Lightfoot E, McCleary JS, Lum T (2014) Asset Mapping as a Research Tool for Community-Based Participatory Research in Social Work. Soc Work Res 38:59–64. https://doi.org/10.1093/swr/svu001 APPENDIX A. System Access and Resources DREAM KG is publicly accessible and fully open-source to support replication, extension, and community-engaged deployment: • Liv...

    work page doi:10.1093/swr/svu001 2014