Elamur Research⌕
☰
home / papers / 2026 / Spatial-RAG: Spatial Retrieval Augmented Generation for Real-World Geospatial Reasoning Questions

Spatial-RAG: Spatial Retrieval Augmented Generation for Real-World Geospatial Reasoning Questions

Paper · 2026 · Dazhou Yu, Riyang Bao, Ruiyu Ning et al.

Summary

Real-world geospatial questions, such as finding a restaurant along a walking route or amenities near a landmark, require both geometric computation and an understanding of what the user wants. LLMs handle language well but cannot do spatial computation and lack current map data. GIS and spatial databases compute geometry precisely but cannot read free-form questions. Spatial-RAG connects the two.

The framework works in three stages:

  1. It translates a natural-language question into spatial SQL, which retrieves objects that satisfy topological, directional or distance constraints, including searches along polylines.
  2. It scores the candidates with a hybrid retriever that combines sparse spatial scores from the database with dense text-embedding similarity.
  3. It treats final selection as a multi-objective trade-off between spatial and semantic relevance. An LLM then picks the answer from the Pareto-optimal candidates according to the inferred user intent.

On tourism and map QA benchmarks, Spatial-RAG achieves the best top-ranked precision and F1 on most comparisons, though not on every metric. The work shows that LLM answers grounded in explicit spatial retrieval are more reliable than relying on the model's internal geographic knowledge.

Key points

  • Benchmarks: The datasets are TourismQA-Miami, TourismQA-NYC, MapQA-ADJ and MapQA-AME. Baselines include GeoLLM, Naive RAG, Sort-by-Distance (SD) and two methods abbreviated TE and ST. An O3-web baseline was run only on Miami.

  • TourismQA results: Spatial-RAG reaches Precision@1 of 0.5152 on Miami versus 0.3947 for the best baseline (ST). On NYC it reaches 0.5665 versus 0.4923 for Naive RAG. On NYC, ST is slightly better at some cutoffs, such as Precision@5/@10 and Recall@3.

  • MapQA results: On MapQA-AME, Spatial-RAG achieves F1 0.7298 (best baseline ST: 0.5346) and NDCG 0.8719. On MapQA-ADJ, its F1 of 0.5625 leads, but ST has higher precision (0.5600 vs 0.5057).

  • Ablation on TourismQA-NYC: Precision@1 falls from 0.5665 when key parts are removed:

    • sparse spatial retrieval: 0.3807
    • Pareto selection: 0.3874
    • reranking: 0.3932
    • retrieval altogether: 0.3584

    Removing the dense spatial component barely changes the results.

  • Limitations: The method needs existing spatial databases with accurate object footprints. The authors also note that recent agentic systems were not systematically evaluated.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 worksSignal Processing · 452 works

🧭 Topics 2Research subjects and languages the work is about.Reasoning · 1,426 worksInformation retrieval & search · 473 works

Cite

Cite as Dazhou Yu, Riyang Bao, Ruiyu Ning, Jinghong Peng, Gengchen Mai and Liang Zhao. 2026. Spatial-RAG: Spatial Retrieval Augmented Generation for Real-World Geospatial Reasoning Questions. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.539

@inproceedings{yu2026spatial,
  title = {{Spatial-RAG: Spatial Retrieval Augmented Generation for Real-World Geospatial Reasoning Questions}},
  author = {Dazhou Yu, Riyang Bao, Ruiyu Ning, Jinghong Peng, Gengchen Mai and Liang Zhao},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.539},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.539}
}

References

30 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Tolgahan Cakaloglu, Christian Szegedy, and Xiaowei Xu. 2020. Text embeddings for retrieval from a large knowledge base. In International Conference on Research Challenges in Information Science, pages 338–351. Springer. find
  2. Wei Chen. 2014. Parameterized spatial sql translation for geographic question answering. In 2014 IEEE international conference on semantic computing, pages 23–27. IEEE. find
  3. Danish Contractor, Shashank Goel, Mausam, and Parag Singla. 2021a. Joint spatio-textual reasoning for answering tourism questions. In Proceedings of the Web Conference 2021, pages 1978–1989. find
  4. Danish Contractor, Barun Patra, Parag Singla, and 1 others. 2021b. Constrained bert bilstm crf for understanding multi-sentence entity-seeking questions. Natural Language Engineering, 27(1):65–87. find
  5. Danish Contractor, Krunal Shah, Aditi Partap, Parag Singla, and 1 others. 2019. Large scale question answering using tourism data. arXiv preprint arXiv:1909.03527. — [arXiv:1909.03527](https://arxiv.org/abs/1909.03527) find
  6. Fahim Faisal and Antonios Anastasopoulos. 2023. Geographic and geopolitical biases of language models. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 139–163. find
  7. Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6491– 6501.
  8. Wes Gurnee and Max Tegmark. 2024. Language models represent space and time. In The Twelfth International Conference on Learning Representations. find
  9. Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. 2024. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. arXiv preprint arXiv:2402.07630.
  10. Xu Chen, Fangning Ren, Dazhou Yu, Lechen Dong, and Fang Liu. 2025. Autosolvate-agent: An autonomous agent for gpu-accelerated solution-phase quantum chemistry. In Large Language Models for Scientific and Societal Advances. find
  11. Yuhan Ji and Song Gao. 2023. Evaluating the effectiveness of large language models in representing textual descriptions of geometry and spatial relations (short paper). In 12th International Conference on Geographic Information Science (GIScience 2023), pages 43–1. Schloss Dagstuhl–Leibniz-Zentrum f… find
  12. Philipp Christmann and Gerhard Weikum. 2024. Rag-based question answering over heterogeneous data and text. arXiv preprint arXiv:2412.07420. — [arXiv:2412.07420](https://arxiv.org/abs/2412.07420) find
  13. Sergios-Anestis Kefalidis, Dharmen Punjani, Eleni Tsalapati, Konstantinos Plas, Maria-Aggeliki Pollali, Pierre Maret, and Manolis Koubarakis. 2024. find
  14. The question answering system geoqa2 and a new benchmark for its evaluation. International Journal of Applied Earth Observation and Geoinformation, 134:104203. adaptation of retrieval augmented generation (rag) models for open domain question answering. Transactions of the Association for Computatio… — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
  15. Nemin Wu, Qian Cao, Zhangyu Wang, Zeping Liu, Yanlin Qi, Jielu Zhang, Joshua Ni, Xiaobai Yao, Hongxu Ma, Lan Mu, and 1 others. 2024. Torchspa-tial: A location encoding framework and benchmark for spatial representation learning. arXiv preprint arXiv:2406.15658. find
  16. Fangjun Li, David C Hogg, and Anthony G Cohn. 2024. Advancing spatial reasoning in large language models: An in-depth evaluation and enhancement using the stepgame benchmark. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18500–18507. find
  17. Zekun Li, Malcolm Grossman, Mihir Kulkarni, Muhao Chen, Yao-Yi Chiang, and 1 others. 2025. Mapqa: Open-domain geospatial question answering on map data. arXiv preprint arXiv:2503.07871. — [arXiv:2503.07871](https://arxiv.org/abs/2503.07871) find
  18. Zekun Li, Wenxuan Zhou, Yao-Yi Chiang, and Muhao Chen. 2023. Geolm: Empowering language models for geospatially grounded language understanding. arXiv preprint arXiv:2310.14478. find
  19. Bastien Liétard, Mostafa Abdou, and Anders Søgaard. 2021. Do language models know the way to rome? arXiv preprint arXiv:2109.07971. find
  20. Gengchen Mai, Weiming Huang, Jin Sun, Suhang Song, Deepak Mishra, Ninghao Liu, Song Gao, Tianming Liu, Gao Cong, Yingjie Hu, and 1 others. 2024. On the opportunities and challenges of foundation models for geoai (vision paper). ACM Transactions on Spatial Algorithms and Systems. find
  21. Gengchen Mai, Krzysztof Janowicz, Rui Zhu, Ling Cai, and Ni Lao. 2021. Geographic question answering: challenges, uniqueness, classification, and future directions. AGILE: GIScience series, 2:8. find
  22. Rohin Manvi, Samar Khanna, Marshall Burke, David B Lobell, and Stefano Ermon. 2024a. Large language models are geographically biased. In International Conference on Machine Learning, pages 34654– 34669. PMLR. find
  23. Rohin Manvi, Samar Khanna, Gengchen Mai, Marshall Burke, David B Lobell, and Stefano Ermon. 2024b. Geollm: Extracting geospatial knowledge from large language models. In The Twelfth International Conference on Learning Representations. find
  24. Jonathan Roberts, Timo Lüddecke, Sowmen Das, Kai Han, and Samuel Albanie. 2023. Gpt4geo: How a language model sees the world’s geography. arXiv preprint arXiv:2306.00020. find
  25. Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen, Tharindu Kaluarachchi, Rajib Rana, and Suranga Nanayakkara. 2023. Improving the domain Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024. Travelplanner: A benchmark for real-world planning… find
  26. Dazhou Yu, Xiaoyun Gong, Yun Li, Meikang Qiu, and Liang Zhao. 2024a. Self-consistent deep geometric learning for heterogeneous multi-source spatial point data prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4001–4011. find
  27. Dazhou Yu, Yuntong Hu, Yun Li, and Liang Zhao. 2024b. Polygongnn: Representation learning for polygonal geometries with heterogeneous visibility graph. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4012–4022. find
  28. Dazhou Yu, Genpei Zhang, and Liang Zhao. 2025. Poly-hedronnet: Representation learning for polyhedra with surface-attributed graph. In The Thirteenth International Conference on Learning Representations. find
  29. Jiong Yu, Sixing Wu, Jiahao Chen, and Wei Zhou. 2024c. Llms as collaborator: Demands-guided collaborative retrieval-augmented generation for commonsense knowledge-grounded open-domain dialogue systems. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13586–13612. find
  30. Yifan Zhang, Zhiyun Wang, Zhengting He, Jingxuan Li, Gengchen Mai, Jianfeng Lin, Cheng Wei, and Wenhao Yu. 2024. Bb-geogpt: A framework for learning a large language model for geographic information science. Information Processing & Management, 61(5):103808. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yu-2026-spatial-rag-spatial-retrieval-augmented/