Neural SQL Engine

Speak. Query.Discover.

The next generation of database interaction. Natural language meets intelligent query generation.

Talk to your databases. Get trusted answers.

Talk to us

What it does

Connect read-only

Connects read-only to the databases you already have. *

Govern the meaning

Builds a governed semantic metadata layer: your tables, your business terms, your rules, curated and versioned.

Ask where work happens

Answers questions where your team already works: Slack, your internal chat portal, and AI assistants like Claude and ChatGPT.

* Initial release supports PostgreSQL only.

Proven on the hardest benchmark

Trusted answers need more than a generated query.

On BIRD, the hardest public text-to-SQL benchmark, DataParlez answers 54 percent of questions exactly right, with 99.8 percent of queries executing. GPT-4 scores 36 percent on the same benchmark, and the best published method 51 percent.1

Our semantic metadata layer gives open-source AI models 1.9 to 2.7 times more correct answers, and lifts every model tested by 16.7 to 21.0 percentage points.2

Our verification pipeline raises the rate of queries that run successfully from 95.1 percent to 99.8 percent.3

54%exact answers

+90%on internal benchmarks4

DataParlez54.2%

GPT-4o + table augmentation50.8%

GPT-435.8%

1. BIRD Mini-Dev, PostgreSQL, 10 databases, 465 held-out questions, pre-registered run, strict result-set match (54.2 percent). Baselines from the BIRD Mini-Dev PostgreSQL leaderboard: GPT-4 35.8 percent, GPT-4o with table augmentation 50.8 percent. Question subsets differ slightly; ours excludes the one database used for development. Benchmark results; customer results depend on data and metadata depth.

2. Peer-reviewed study, BIRD Mini-Dev, 11 databases, six models, 21,000 evaluated questions, statistically significant for every model. Multipliers versus schema-only prompting: Qwen3.5-9B 10.3 to 28.2 percent, Qwen3.5-27B 14.7 to 35.6, Qwen3.6-35B 17.4 to 37.3, Gemma-4-26B 15.1 to 29.0. Claude Opus 4.7: 32.2 to 44.4.

3. Pre-registered test on 465 held-out benchmark questions across 10 databases. Execution rate measures queries that run and return results; it is a reliability measure, not answer correctness.

4. Internal DataParlez benchmarks. Results depend on the database, questions, model, and metadata depth and are not directly comparable to BIRD.

In production

Built for the environment you already trust.

Connect DataParlez to Slack, Microsoft Teams, LibreChat, Open WebUI, Gemini, Claude, and ChatGPT.

Your data stays yours

Runs entirely inside your cloud, with your own AI provider keys. Your data never leaves your tenant, and the system can be locked to read-only access.

For full sovereignty, contact us.

Runs in your cloud

Your AI provider keys

Read-only access

Single-tenant deployment

Abstract private cloud deployment boundary

Start with a pilot

90-day pilot on your own data. We set up the metadata layer with your team, agree success criteria up front, and you decide at day 75.

75You decide before the pilot ends.

About DataParlez

DataParlez makes databases answer questions in plain language, reliably.

We build DataParlez on a Neural SQL engine that combines deterministic and agentic workflows with a governed semantic metadata layer and a verification pipeline that checks every query before it runs.