SANA

How SANA works

When an LLM agent fails at Exploratory QA over a data lake, which part of its runtime is to blame? An agent answers in three stages. SANA replaces one stage at a time with an oracle built from the task's ground truth — the accuracy it gains pinpoints the bottleneck.

1. 1

   Planning

   Breaks the question into an ordered list of sub-questions.

   oracle hands over the correct sub-questions directly.

2. 2

   Search

   Hunts a ~40M-file lake for the datasets each sub-question needs.

   oracle returns only the datasets the task actually requires.

3. 3

   Data Analysis

   Writes and runs SQL/Python to compute each intermediate answer.

   oracle executes the agent's intent with zero implementation bugs.

Swap any oracle for your own implementation and watch the delta move — see the Pipeline.

---

The paper link is a placeholder pending the arXiv release.

---

Citation

To appear at the VLDB 2026 Workshop on Systems for Data-centric Agents with Human-in-the-loop (DASHSys).

@inproceedings{sana2026,
  title     = {SANA: What Matters for QA Agents over Massive Data Lakes?},
  author    = {Wijaya, Austin Senna and Liu, Jiaxiang and Wang, Haonan and Wu, Eugene},
  booktitle = {VLDB 2026 Workshop on Systems for Data-centric Agents with Human-in-the-loop (DASHSys)},
  year      = {2026}
}