live wire
▸nl2sql.ai is live — first leaderboard edition published, directory and benchmark tracker onlinenl2sql.ai▸BIRD leaderboard: top cluster holds in the low-to-mid 70s on dev execution accuracy; human reference 92.96bird-bench.github.io▸Spider 2.0 remains the wall: launch-paper agentic baseline solved ~17% of enterprise tasksarXiv▸Directory day one: 12 systems catalogued across cloud-native, OSS, and research categoriesnl2sql.ai▸Vanna remains the most-starred OSS text-to-SQL framework; RAG-on-your-own-pairs still the default patternGitHub▸Wren AI ships steadily on its MDL semantic layer — the OSS counterpart to vendor semantic modelsGitHub▸Uber QueryGPT post remains the canonical enterprise-scale case study: routing beats generation at 1000s of tablesUber Engineering▸Watch item: semantic-layer interop — every platform has one, none of them talk to each otheranalysis desk▸DB-GPT community keeps shipping: fine-tuning hub and AWEL workflows anchor the self-hosted stackGitHub▸Desk assignments filed: releases and papers, leaderboards, the directory. Cadence: continuousnl2sql.ai▸nl2sql.ai is live — first leaderboard edition published, directory and benchmark tracker onlinenl2sql.ai▸BIRD leaderboard: top cluster holds in the low-to-mid 70s on dev execution accuracy; human reference 92.96bird-bench.github.io▸Spider 2.0 remains the wall: launch-paper agentic baseline solved ~17% of enterprise tasksarXiv▸Directory day one: 12 systems catalogued across cloud-native, OSS, and research categoriesnl2sql.ai▸Vanna remains the most-starred OSS text-to-SQL framework; RAG-on-your-own-pairs still the default patternGitHub▸Wren AI ships steadily on its MDL semantic layer — the OSS counterpart to vendor semantic modelsGitHub▸Uber QueryGPT post remains the canonical enterprise-scale case study: routing beats generation at 1000s of tablesUber Engineering▸Watch item: semantic-layer interop — every platform has one, none of them talk to each otheranalysis desk▸DB-GPT community keeps shipping: fine-tuning hub and AWEL workflows anchor the self-hosted stackGitHub▸Desk assignments filed: releases and papers, leaderboards, the directory. Cadence: continuousnl2sql.ai
nl2sql.ai
analysisTHE LANDSCAPE

The state of natural-language-to-SQL, August 2026

Semantic layers won the argument, agents won the benchmarks, and the hard problem moved from writing SQL to trusting it.

By The Editorial Desk· Aug 10, 2026

Three shifts define where natural-language-to-SQL stands today.

1. Semantic grounding beat raw generation

The first generation of NL2SQL — prompt a model with a schema dump and hope — died on contact with real warehouses. What replaced it is governed context: Snowflake's Cortex Analyst grounds generation in a semantic model of measures, dimensions, and verified queries; Databricks Genie scopes conversations to curated spaces; open-source Wren AI builds an explicit modeling layer (MDL) before any SQL is written. The pattern is the same everywhere: the model doesn't guess what "revenue" means — the semantic layer tells it.

The retrieval-first OSS stack made the same move from the other direction. Vanna's train-on-your-pairs loop and Dataherald's context stores are semantic layers by another name: curated, versioned context that stands between the user's question and the model's SQL.

2. Single queries became agentic workflows

Benchmarks tell the story. Yale's Spider (2018) was effectively solved — top systems above 91% — so the field built harder tests. BIRD (2023) added dirty values, external knowledge, and 33 GB databases; the best systems now cluster in the low-to-mid 70s on execution accuracy against a 92.96% human reference. Spider 2.0 (late 2024) rebuilt the task around real enterprise projects — thousand-column Snowflake and BigQuery schemas, multiple dialects, multi-step workflows — and frontier models landed under 20% at launch.

The systems that climb these leaderboards are not single-shot generators. They explore schemas, draft candidate queries, execute, inspect failures, and revise — small agentic pipelines, exactly the architecture Uber described for its internal QueryGPT.

3. The product problem is now trust

Accuracy improved enough that the binding constraint moved: the question enterprises ask is no longer "can it write SQL?" but "how do I know this particular answer is right?" Hence verified-query workflows, feedback loops, audit trails, and permissions-aware generation showing up as headline features across every platform. The vendors that treat wrong-but-plausible answers as the core design problem — not an edge case — are the ones winning enterprise deployments.

What we're watching

  • Whether Spider 2.0 scores move the way BIRD's did — and which architectures move them.
  • Open-weight, SQL-specialized models (the SQLCoder lineage) versus frontier generalists on cost-per-correct-query.
  • Semantic-layer standardization: every vendor has one; none of them interoperate.
  • NL2SQL disappearing as a product category — absorbed into "analyst agents" that plan, query, chart, and narrate.

This piece anchors our coverage; the leaderboard and the wire carry it forward from here.

sources

Filed by The Editorial Desk. Corrections: desk@nl2sql.ai · Our standards →
more from the desk