ProSearch: Benchmarking Multi-Constraint Protocol Retrieval in Experimental Science

ProSearch is a benchmark for retrieving scientific method paragraphs that satisfy natural-language queries containing multiple experimental constraints. A relevant paragraph should support the constraints within one coherent experimental setup.

The benchmark covers six experimental-science domains, with 1,515 queries and 358,178 method paragraphs.

Dataset structure

Each domain directory contains four JSON files:

prosearch/
├── cell_culture/
├── electrochemistry/
├── immunotechnology/
├── inorganic_materials/
├── organic_synthesis/
└── polymer_materials/

Each directory contains:

File	Description
paragraphs.json	Method paragraph candidate corpus used in the main experiments reported in the paper.
all_paragraphs.json	Full paragraph candidate corpus used for full-text retrieval comparisons.
queries_raw.json	Raw queries for the domain.
qrels_stage1.json	Query–paragraph relevance judgments. The JSON object maps each query ID to paragraph IDs with relevance label 1.

Dataset statistics

Domain	Queries	Method paragraph pool	Full paragraph pool
Cell culture	284	46,004	311,042
Electrochemistry	201	91,375	409,133
Immunotechnology	264	93,315	644,077
Inorganic materials	247	46,136	240,380
Organic synthesis	252	30,371	197,506
Polymer materials	267	50,977	283,907
Total	1,515	358,178	2,086,045

The method paragraph pool is stored in paragraphs.json, while the full paragraph pool is stored in all_paragraphs.json. The method paragraph pool is used for the main benchmark experiments. The full paragraph pool is used for the full-text retrieval analysis and contains paragraphs from all body sections of the paper. Only the candidate pool changes between the two settings; the queries and relevance definition remain the same.

Data construction

The data were derived from English-language open-access scientific articles. The construction pipeline includes article parsing, method-paragraph selection, structured protocol extraction and normalization, query construction, candidate-pool construction, and relevance judgment against the original paragraph text.

Candidate-pool membership alone does not determine relevance; relevance labels are provided in the corresponding qrels files.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. See the license text.

Citation

If you use ProSearch, please cite:

ProSearch: Benchmarking Multi-Constraint Protocol Retrieval in Experimental Science. NeurIPS 2026 Evaluations & Datasets Track.

Limitations

While ProSearch provides a systematic evaluation of multi-constraint protocol retrieval, several limitations remain. First, the field schemas and constraint rules are designed by domain experts for selected experimental protocol types across six domains, limiting coverage of other experimental procedures and requiring additional expert effort for extension. Second, ProSearch focuses on individual method paragraphs and does not evaluate evidence integration across multiple paragraphs, tables, or supplementary materials. Third, LLM-based protocol extraction and relevance labeling, together with annotation pools constructed from extracted records, may introduce errors or omit relevant evidence. Although expert audits and robustness analyses support the stability of the main retrieval findings, systematic biases cannot be fully ruled out. Finally, while the queries are grounded in real experimental records, they are generated using templates and may not fully reflect researchers' actual information needs. Whether benchmark scores translate into retrieval usefulness in real scientific workflows also remains to be established. Future work will extend ProSearch to cross-paragraph protocol retrieval and incorporate researcher-authored queries and human evaluations to assess its practical applicability.