SQLord: Enterprise NL2SQL Framework

Updated 6 July 2026

SQLord is an enterprise-level NL2SQL framework that translates natural language into complex SQL, using reverse data generation for robust domain adaptation.
It employs a multi-step workflow with task decomposition and GPT-Judge evaluation to ensure semantically correct queries even when standard benchmarks are inadequate.
Benchmark results and ablation studies show SQLord achieves over 86% execution accuracy, highlighting its effectiveness in complex production environments.

SQLord is an enterprise-level natural-language-to-SQL framework introduced as a robust text-to-SQL solution via reverse data generation and workflow decomposition, intended for environments in which complex business logic, scarce labeled NL2SQL data, and limited access to executable ground-truth databases make benchmark-style systems inadequate (Cheng et al., 14 Jul 2025). In later technical syntheses, the label is also used more broadly for systems that verify, orchestrate, or semantically mediate SQL generation, so the term denotes both a specific Alibaba deployment and a wider design pattern for SQL-centered reasoning over complex data systems (Tritto et al., 1 Sep 2025, Kim et al., 30 Jun 2026).

1. Origins, antecedents, and problem setting

The immediate problem addressed by SQLord is enterprise NL2SQL. In business environments, queries often encode domain-specific notions such as “best-selling product,” “competitor,” and “customs import–export status,” while production SQL may require many-table joins, nested subqueries, complex filters, case logic, aggregations, and multiple grouping levels. Enterprise logs contain large numbers of SQL queries written by engineers or BI specialists, but almost none are paired with the natural-language questions that motivated them. Evaluation is also constrained: in production, one may have only natural-language queries and generated SQL, no gold SQL, or gold SQL without a convenient executable sandbox (Cheng et al., 14 Jul 2025).

Earlier systems addressed neighboring parts of the same problem space. SODA offered a Google-like search experience for data warehouses by taking keyword queries of business users and automatically generating executable SQL through graph pattern matching over a metadata model of the data warehouse (Blunschi et al., 2012). Cyrus provided a mobile, voice-driven interface to arbitrary relational databases for tutoring and formative assessment of SQL learners, with schema-aware matching, string similarity, and result-equivalence grading (Godinez et al., 2018). These systems did not formulate the same three-part enterprise agenda as SQLord, but they established recurrent themes: metadata-grounded interpretation, user-facing query mediation, and semantic rather than purely syntactic correctness.

A plausible implication is that SQLord emerged where these lines of work intersected with LLM-era requirements: domain adaptation from unlabeled SQL logs, adaptive multi-step query generation, and evaluation methods that remain usable when gold labels or execution environments are missing.

2. Enterprise framework and system architecture

SQLord is organized around three pillars: Reverse Data Generation, Automated Workflow (Task) Decomposition, and a GPT-Judge Evaluation Framework (Cheng et al., 14 Jul 2025). The framework is described as follows: SQL logs plus a small set of <SQL, COMMENT> pairs are used to train a reverse model; the reverse model generates <Query, SQL> pairs; a domain NL2SQL model is fine-tuned on these pairs; a retrieval-augmented workflow generator decomposes complex user questions into sub-tasks; and generated SQL is evaluated with EXE, QSE, and SSE.

Component	Role	Model
RevLLM	SQL to natural-language description	Qwen2-7B-Instruct
SQLLM	Domain NL2SQL model	Qwen2-7B-Instruct
GPT-Judge	EXE, QSE, and SSE evaluation	GPT-4

RevLLM is trained to map SQL to comment-style natural-language descriptions. SQLLM is then fine-tuned on reverse-generated <Query, SQL> data so that the model is adapted to enterprise schemas and business terminology rather than only to open benchmarks. GPT-Judge functions as an evaluation layer for settings with or without gold SQL and with or without an executable database.

The architecture is explicitly aimed at bridging the research–production gap. Existing LLM-based NL2SQL systems are described as largely optimized for static benchmarks like Spider, with heavily engineered prompts and in-context examples, but not designed around the realities of enterprise SQL logs, evolving schemas, business-specific logic, and partial observability during evaluation (Cheng et al., 14 Jul 2025).

3. Reverse data generation and supervised adaptation

The reverse data generation method begins with a seed annotated dataset of 73,589 <SQL, COMMENT> pairs from routine development, where each SQL statement is paired with a human-written comment describing what the SQL does (Cheng et al., 14 Jul 2025). RevLLM is trained on this task and then applied to a large SQL log set to generate a reverse-generated training corpus of 35,948 <Query, SQL> pairs. This pseudo-annotated corpus is used to fine-tune SQLLM.

The training setup is concrete. The base model is Qwen2-7B-Instruct. SQLLM is trained for 3 epochs with learning rate 1e-5, batch size 1, on 8 × NVIDIA H100 GPUs (Cheng et al., 14 Jul 2025). The purpose of this supervised fine-tuning is domain adaptation: the model learns real internal table and column names, organization-specific business conventions, and recurrent analytical query forms that do not appear in open academic benchmarks.

The paper formalizes the downstream objective as autoregressive SQL generation conditioned on the query. In prose, SQLLM minimizes the negative log-likelihood of SQL tokens given the natural-language query and the SQL prefix. The significance of this step is not merely synthetic data augmentation. Because the reverse-generated pairs are drawn from production SQL logs, the resulting model inherits the organization’s existing analytical idioms, including join patterns, metric definitions, and business filters.

The reported ablation shows why this matters. On the real-world dataset, base Qwen reaches 51.3% EXE, while + Reverse Generation reaches 63.1% EXE; adding workflow generation then raises EXE to 74.6%, and the full SQLord system reaches 86.5% (Cheng et al., 14 Jul 2025). This establishes reverse generation as a primary mechanism for converting unlabeled enterprise SQL into usable supervision.

4. Workflow decomposition and GPT-Judge evaluation

SQLord treats complex NL2SQL not as one monolithic decoding step but as a workflow of sub-tasks. Given a user query $q$ , the system first retrieves relevant domain knowledge and table schemas,

$d_q = \text{RA}(q; \mathcal{K}, \mathcal{T}),$

where $\mathcal{K}$ is a domain knowledge base and $\mathcal{T}$ is a schema repository called TableHub (Cheng et al., 14 Jul 2025). It then generates sub-tasks step by step,

$t_i = G(q, d_q, \{r_1, \dots, r_{i-1}\}),$

so that each sub-task can depend on prior execution results. For each sub-task, SQLLM generates SQL and an execution engine returns intermediate results,

$s_i = \text{SQLLM}(t_i, d_q; \phi), \quad r_i = \mathcal{E}(s_i).$

Final outputs are aggregated as

$r_q = \text{Summary}(\{r_1, \dots, r_k\}).$

This decomposition is retrieval-augmented and adaptive. Earlier steps may identify relevant tables or metrics, whereas later steps may refine filters or compute derived metrics based on previously returned results. The design directly targets enterprise questions that mix lookup, aggregation, comparison, and business interpretation.

SQLord’s evaluation framework contains three modes. When both gold SQL and an executable database are available, Execution Evaluation (EXE) compares execution results:

$\text{EXE Score} = \mathbb{I}(r_{\text{gen}} = r_{\text{gold}}).$

When gold SQL is unavailable, Query-SQL Evaluation (QSE) asks an LLM judge to assess whether generated SQL implements the natural-language request given schema and domain context. When gold SQL exists but execution is unavailable or inconvenient, SQL-SQL Evaluation (SSE) asks an LLM judge to compare generated SQL and reference SQL semantically (Cheng et al., 14 Jul 2025).

The framework is designed for practical production combinations of observability and supervision. EXE is standard execution accuracy. QSE is useful for unlabeled online traffic. SSE is useful when SQL is available but execution is not. The paper reports “a consistent gap between EXE and GPT-Judge (±7% for QSE and ±1% for SSE), validating its effectiveness as an execution-independent evaluation method” (Cheng et al., 14 Jul 2025).

5. Benchmarks, ablations, and production deployment

SQLord is evaluated on a real-world enterprise dataset of 6,000 manually annotated <Query, SQL> pairs drawn from two online scenarios, Customs Import–Export Assistant and Intelligent Product Selection, and also on Spider (Cheng et al., 14 Jul 2025). On the real-world dataset, the reported scores are: Qwen 51.3 / 58.1 / 51.7 for EXE / QSE / SSE, GPT-4 68.5 / 75.6 / 69.2, DIN-SQL (GPT) 71.4 / 78.3 / 71.9, DAIL-SQL (GPT) 72.9 / 78.7 / 72.3, and SQLord (Qwen) 86.5 / 93.2 / 85.8. On Spider, the corresponding scores are: Qwen 66.7 / 72.9 / 67.1, GPT-4 80.2 / 87.6 / 80.8, DIN-SQL (GPT) 85.3 / 91.5 / 85.9, DAIL-SQL (GPT) 86.6 / 93.4 / 86.8, and SQLord (Qwen) 87.4 / 93.5 / 87.2 (Cheng et al., 14 Jul 2025).

The ablation study attributes these gains to cumulative architecture rather than to one isolated component. On the real-world dataset, Qwen achieves 51.3 / 58.1 / 51.7, + Reverse Generation achieves 63.1 / 69.8 / 63.7, + Workflow Generation reaches 74.6 / 81.2 / 74.3, and the full system reaches 86.5 / 93.2 / 85.8 (Cheng et al., 14 Jul 2025). The progression indicates that SQLord’s contribution lies in the combination of synthetic supervision, decomposition, and evaluation.

Deployment results are reported on the world’s largest B2B e-commerce platform in the two named scenarios. Relative to the pre-SQLord system, Customs Import–Export Assistant improves by +15.6 EXE, +14.4 QSE, and +16.1 SSE, while Intelligent Product Selection improves by +16.4 EXE, +15.9 QSE, and +16.7 SSE (Cheng et al., 14 Jul 2025). The abstract states that “online accuracy consistently exceeds 90,” and the system is described as successfully applied across multiple scenarios on that platform.

These results position SQLord as a production NL2SQL framework rather than a benchmark-only model. Its deployment pattern combines domain retrieval, workflow generation, SQL execution over business warehouses, and optional summarization of final results.

In later research syntheses, “SQLord” is used in a broader, interpretive sense for systems that govern SQL generation by ranking, semantic mediation, or orchestration. One such interpretation appears in GradeSQL, where “a ‘SQLord’ system, in the sense of this paper, is exactly what GradeSQL builds”: a principled verifier that scores and ranks many LLM-generated SQL candidates, then selects the most semantically correct one (Tritto et al., 1 Sep 2025). GradeSQL defines an Outcome Reward Model with candidate set $Y_{\text{candidate}} = \{c_1, \dots, c_N\}$ , scoring function $\phi(q_i, c_j) \in [0,1]$ , and ranked choice $d_q = \text{RA}(q; \mathcal{K}, \mathcal{T}),$ 0. On N=32 candidates from OmniSQL-7B, ORM-BoN reaches 68.90% on BIRD dev, 84.53% on Spider dev, and 87.47% on Spider test, outperforming both execution-based BoN and Majority Voting (Tritto et al., 1 Sep 2025).

A second extension appears in the semantic-layer-mediated NL2SQL agent literature. A “SQLord”-style system is described there as one that decouples semantic intent from physical SQL execution through a curated semantic layer and an intermediate representation called the Semantic Model Query (SMQ) (Kim et al., 30 Jun 2026). That agent supports SQLite, BigQuery, and Snowflake backends and reaches 94.15% execution accuracy on the 547-task Spider2-snow benchmark, ranking third on the official leaderboard (Kim et al., 30 Jun 2026). The commonality with SQLord is the emphasis on explicit orchestration and grounding rather than direct raw-schema decoding.

Step-wise orchestration is pursued explicitly in SQLConductor, which formulates Text-to-SQL subtasks as specialized actions and learns a policy that chooses the next action based on intermediate artifacts and feedback. With a compact policy coordinating frozen larger action models, it reaches 73.2% EX on BIRD-Dev (Zhu et al., 22 Jun 2026). THOR provides a production-oriented complement: a decoupled orchestration/execution architecture with on-demand schema injection, a SQL Generation Agent that emits single-statement SELECT queries protected by a read-only guardrail, and a self-correction loop with up to five LLM-driven regeneration attempts (Shi et al., 13 Jul 2025). These systems do not share SQLord’s exact three-pillar design, but they inhabit the same design space of adaptive SQL mediation.

Other adjacent lines of work broaden the conceptual perimeter further. SODA used graph pattern matching over a metadata model to produce SQL for business users in complex warehouses (Blunschi et al., 2012). Cyrus provided voice-first NL-to-SQL tutoring with result-equivalence grading and schema-aware matching (Godinez et al., 2018). SABER proposed a semantic algebra based on extended relational algebra for SQL-compatible semantic document processing (Lee et al., 29 Aug 2025). iPDB integrated ML and LLM inferencing into SQL through a relational predict operator $d_q = \text{RA}(q; \mathcal{K}, \mathcal{T}),$ 1 and semantic query optimizations (Kumarasinghe et al., 23 Jan 2026). A plausible implication is that “SQLord” now names not only one enterprise framework but also a recurrent research objective: to place a learned, semantically informed control layer over SQL generation, execution, and interpretation.