For enterprises relying on NoSQL databases as their core data infrastructure, the ability to query them using natural language remains underdeveloped. A new research paper from arXiv presents TEND (Text-to-NoSQL Dataset) and SAG (Schema-as-Data Grounding), aiming to bridge this gap for MongoDB aggregation pipelines over schema-less document stores.
According to the paper, correct query generation must recover how a non-relational data model represents entities, nested paths, arrays, missing fields, and dynamic keys. This challenge is more complex than traditional SQL querying because NoSQL databases like MongoDB store data without a fixed schema.
The Challenge of Schema-less Document Stores
NoSQL databases are widely used for their flexibility, but natural-language access to them remains underdeveloped. The authors, including Lu, Jinwei, Jiawei, Zhang, Chen, Qin, Zhiqian, Haodi, Song, Yuanfeng, Wong, and Raymond Chi-Wing, note that translating natural language requests into executable NoSQL queries requires understanding how entities and relationships are encoded in non-relational models. For example, a query must handle nested arrays, optional and sparse paths, and polymorphic shapes—features not present in relational databases.
TEND: An Execution-Verified Benchmark
The paper presents TEND, an execution-verified benchmark with 1,210 MongoDB-native tasks across 11 databases. To the authors' knowledge, TEND is the first Text-to-NoSQL benchmark whose database worlds are MongoDB-native by design. Experts manually defined collection boundaries, nested arrays, optional and sparse paths, polymorphic shapes, and dynamic-key conventions. The worlds are populated with real data and verified through frozen MongoDB execution. This ensures that TEND evaluates schema-less document reasoning rather than SQL-to-MQL transfer.
SAG: Schema-as-Data Grounding Solver
The authors further introduce SAG, a Schema-as-Data Grounding solver. SAG induces path and value grounding from stored-document evidence before bounded MQL generation, followed by execution-grounded repair and result-consistency selection. Evaluation uses bounded column-tolerant execution accuracy (EXC) as the headline metric, complemented by a graded result-set F1 and a mutually exclusive execution-outcome decomposition.
Implications for AI-Powered Data Access
Experiments demonstrate that large language models (LLMs) with strong NL2SQL performance degrade substantially on TEND, validating Text-to-NoSQL as a distinct schema-less document reasoning problem. This finding highlights the need for specialized approaches when applying natural language interfaces to NoSQL databases. For enterprises, this research points to a future where complex querying of diverse data stores becomes more accessible, but significant work remains to match the maturity of SQL-based solutions.
| Aspect | NL2SQL (Relational) | Text-to-NoSQL (Document) |
|---|---|---|
| Schema | Fixed, known schema | Schema-less, dynamic keys |
| Data model | Tables and joins | Nested arrays, optional paths |
| Query generation | Mature benchmarks | First benchmark (TEND) |
| LLM performance | Strong | Substantially degrades |
The paper is available on arXiv and represents a foundational step for enabling natural language querying of NoSQL systems, a critical capability for data-driven enterprises managing diverse and flexible data architectures.
