pxyu/SQL-Schema-Retrieval

SQL Schema Retrieval

The evaluation benchmark from "Finding the Right Tables and Columns: A Benchmark and Corpus-Adaptive Embeddings for SQL Schema Retrieval" (Zeng, Yu, Mehta, Zhao, Samdani).

Task — schema linking as retrieval. Given a natural-language question over a database, retrieve the schema element(s) needed to answer it. Documents are schema elements: at table granularity, each document is a table schema (columns + types rendered as markdown, with sample rows); at column granularity, each document is a single column with its table context. Relevance R(q) = the schema elements referenced by the question's gold SQL (obtained by parsing table/column references from the ground-truth query). Retrieval is performed per database group (rank the schema at hand). Metrics: nDCG@10 and recall@10.

This recasts five text-to-SQL datasets as retrieval, spanning academic → enterprise → large live schemas — a step that pure text-to-SQL accuracy never exercises because it assumes the whole schema fits in context.

Subsets

Counts below are measured from this release (corpus docs = retrieval candidates at the subset's granularity; rel. = relevant query–document judgments).

Subset (this release)	Source	Granularity	Corpus docs	Queries	Rel. judgments
spider	Spider (academic, cross-domain)	table	180	2,147	3,366
bird	BIRD (realistic, value semantics)	table	75	1,534	2,956
beaver	BEAVER (private enterprise warehouses)	table	463	209	928
livesqlbench_table	LiveSQLBench (base)	table	244	410	1,075
livesqlbench_large_table	LiveSQLBench-Large	table	971	332	901
livesqlbench_large_column	LiveSQLBench-Large	column	17,709	332	2,157

Underlying database sizes (from the paper's Table 1): Spider 40 DB / 785 col; BIRD 11 DB / 798 col; BEAVER 6 DB / 4,238 col; LiveSQLBench 22 DB / 1,942 col; LiveSQLBench-Large 18 DB / 17,708 col. Gold sets average 1.6–4.4 tables (table level) and up to 6.5 columns (column level) per query. The paper additionally studies two document representations (schema-metadata vs. value-only) for Spider/BIRD/BEAVER; this release provides the table/column collections used for evaluation.

Format (BEIR / MTEB retrieval layout)

<subset>/
  corpus.jsonl        # {"_id": "<schema-element id>", "title": "", "text": "<schema as markdown>"}
  queries.jsonl       # {"_id": "<qid>", "text": "<natural-language question>"}
  qrels/test.tsv      # query-id \t corpus-id \t score   (relevant judgments, score>0)

_id encodes the schema element as provided by the source: db__table for table-level subsets, and a column identifier (e.g. db__table__column) for the column-level subset.

Sources & citation

This benchmark repackages five public text-to-SQL datasets as retrieval; please cite this work and the original datasets, and follow each source's license/terms:

• Spider — Yu et al., 2018 (EMNLP). https://yale-lily.github.io/spider
• BIRD — Li et al., 2023 (NeurIPS D&B). https://bird-bench.github.io/
• BEAVER — Chen et al., 2024, arXiv:2409.02038.
• LiveSQLBench (base + large) — BIRD Team, 2024. https://github.com/bird-bench/livesqlbench

@inproceedings{zeng2025sqlschemaretrieval,
  title     = {Finding the Right Tables and Columns: A Benchmark and Corpus-Adaptive Embeddings for SQL Schema Retrieval},
  author    = {Zeng, Qingcheng and Yu, Puxuan and Mehta, Aman and Zhao, Fuheng and Samdani, Rajhans},
  year      = {2025},
  note      = {Update venue/URL on publication}
}