Elamur Research⌕
☰
home / papers / 2026 / Routing End User Queries to Enterprise Databases

Routing End User Queries to Enterprise Databases

Paper · 2026 · Saikrishna Sudarshan, Tanay Kulkarni, Manasi Patwardhan et al.

Summary

The paper studies database routing: given a natural-language question and a large set of enterprise databases, rank the databases by how likely each one can answer it. The authors argue that an earlier benchmark (Mandal et al., 2025) kept the original dataset splits, which made test repositories tiny and produced misleading results. They build Spider-Route and Bird-Route by pooling all databases from Spider and BIRD-SQL into one shared repository and splitting each database's questions evenly into train and test. They define in-domain and cross-domain settings over the same repository. Their training-free method first ranks databases by embedding similarity with gte-Qwen2-7B-instruct. It then re-ranks the top candidates with Gemini 2.0 Flash, breaking the task into smaller checks: join adjacency lists, mapping query phrases to schema entities, and checking coverage and connectivity. Each database gets a total score, with a semantic score used to break ties. The method beats embedding-only retrieval and direct LLM re-ranking on all metrics, reaching roughly 80% Recall@1. This shows that structured reasoning helps when databases have overlapping schemas.

Key points

  • Benchmarks:
    • Spider-Route has 206 DBs and 11,831 questions (5,892 train / 5,939 test).
    • Bird-Route has 80 DBs (69 train + 11 dev from BIRD-SQL) with 5,461 train / 5,501 test questions.
    • For Bird-Route, the BIRD "evidence" is aggregated per DB and used as metadata.
  • Pipeline:
    • Cosine similarity between query and DDL-schema embeddings selects the top-k candidates.
    • An LLM then builds join adjacency lists (computed once per DB) and maps query spans to columns and values.
    • Candidates are re-ranked by a total score, and ties are broken by an embedding-based semantic score.
  • Remaining errors (about 20% on Spider-Route, 21% on Bird-Route):
    • In some cases, the correct DB was missing from the embedding top-5.
    • When the correct DB is inserted into the candidates, the re-ranker places it first in 98% (Spider-Route) and 96% (Bird-Route) of those cases.
  • Stronger setup: On a 500-query subset of errors, switching to Gemini 2.5 with top-20 candidates gave Recall@1 of 92% (Spider-Route) and 89% (Bird-Route). Most leftover failures stem from ambiguous queries that more than one DB could legitimately answer.
  • Metadata ablation: Removing domain metadata on Bird-Route lowers Recall@1 by 5% and Recall@3 by 7%.

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

 Institutions

🎓Birla Institute of Technology and Science, PilaniPilāni · India16 works · 41,879 citations🎓Birla Institute of Technology and Science, Pilani - Goa CampusSancoale · India4 works · 1,114 citations🏢Tata Consultancy ServicesPune · India8 works · 35 citations

 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 worksComputer Networks and Communications · 64 works

Cite

Cite as Saikrishna Sudarshan, Tanay Kulkarni, Manasi Patwardhan, Lovekesh Vig, Ashwin Srinivasan and Tanmay Tulsidas Verlekar. 2026. Routing End User Queries to Enterprise Databases. Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.surgellm-1.16

@inproceedings{sudarshan2026routing,
  title = {{Routing End User Queries to Enterprise Databases}},
  author = {Saikrishna Sudarshan, Tanay Kulkarni, Manasi Patwardhan, Lovekesh Vig, Ashwin Srinivasan and Tanmay Tulsidas Verlekar},
  booktitle = {Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.surgellm-1.16},
  url = {https://doi.org/10.18653/v1/2026.surgellm-1.16}
}

References

7 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, and Noveen Sachdeva. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261.
  2. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aieleen Letman, Akhil Mathur, Alan Schelten, and Amy Yang. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  3. Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, Xuanhe Zhou, Chenhao Ma, Guoliang Li, Kevin Chen-Chuan Chang, Fei Li, Reynold Huang, and Zhanwei Lan. 2024. Can LLM already serve as a database interface? a BIg bench for large-scale datab…
  4. Priyangshu Mandal, Manasi Patwardhan, Mayur Patidar, and Lovekesh Vig. 2025. Dbrouting: Routing end user queries to databases for answerability. arXiv preprint arXiv:2501.16220. — [arXiv:2501.16220](https://arxiv.org/abs/2501.16220) find
  5. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992.…
  6. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Juny… find
  7. Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sudarshan-2026-routing-end-user-queries-enterprise/