How Tencent Cloud’s TCDataAgent Cracked the BIRD‑Bench NL2SQL Challenge

On July 4, the international BIRD‑Bench benchmark announced that Tencent Cloud’s self‑built data‑analysis agent TCDataAgent ranked third worldwide, surpassing companies such as IBM and Meta, and achieving the highest domestic ranking on the list.

What is NL2SQL and why is BIRD‑Bench so hard? NL2SQL (Natural Language to SQL) converts everyday natural language into structured SQL queries, enabling “human‑language data retrieval.” BIRD‑Bench evaluates NL2SQL under extremely strict conditions using a massive 33 GB dataset covering 37 industries (finance, healthcare, sports, etc.) with noisy, real‑world data and tens of thousands of complex queries.

//真实场景: 基于金融、医疗、体育等37个行业真实、庞大（33GB）、含“脏数据”的数据库，上万条复杂查询问题。

Beyond understanding natural language, the generated SQL must run efficiently on real databases without errors such as Cartesian products, which can cripple performance.

//能翻译，跑得快: SELECT product_name FROM sales WHERE sale_date BETWEEN '2024-06-01' AND '2024-06-30' ORDER BY sale_amount DESC LIMIT 1;

Under these harsh conditions, TCDataAgent scored 75.74 points.

What makes TCDataAgent stand out?

It introduces a database constraint verification mechanism that automatically detects and corrects structural or semantic SQL errors such as missing join conditions that cause Cartesian products.

It tightly couples SQL generation with the actual database content, optimizing the produced queries to dramatically improve intent recognition accuracy and result reliability.

It employs a post‑training technique that selects the best‑performing SQL samples for iterative model training, enhancing sample quality and model stability.

The research results have been accepted at the top‑tier VLDB conference. Experiments show that the core “database‑content awareness” module can be integrated into other NL2SQL systems, boosting query execution accuracy by up to 18.3% and delivering more than a 5% performance improvement over mainstream methods.

In the era of large models, AI‑driven innovation defines the future of data. TCDataAgent embodies Tencent Cloud’s Data+AI synergy, aiming to let users query complex data through natural language without needing deep database expertise, bringing the vision of “everyone as a data scientist” closer to reality.

—END—