ByteDance's internal data analysis assistant iDA—used inside the company for three years—has officially launched on Volcano Engine. After reading its technical details, our judgment is clear: getting AI to write a working SQL query accounts for only 20% of the work. The remaining 80% is engineering grind—dialect adaptation, business-logic reconciliation, and permissions.

What This Is

iDA is ByteDance's widely used internal data intelligence assistant. Around 2024, its form was still close to ChatBI—letting the model translate natural language into SQL, freeing users from drag-and-drop BI data extraction. After real productization, the team discovered a substantial engineering gap between "can write SQL" and "can reliably query data."

It's now open to external users, with capabilities covering data discovery, data analysis, knowledge retrieval, and document writing. Three key choices hide inside the technical details:

First: DSL or SQL. A DSL (Domain-Specific Language) is a custom syntax—engineering-friendly but unfamiliar to models. iDA chose to let the model write SQL directly, because SQL is abundant in model training corpora. Spending model attention on problem decomposition pays off better.

Second: single Agent or multi-Agent. An Agent is an AI assistant that autonomously completes multi-step tasks. Multi-Agent collaboration is the current trend, but iDA chose a main Agent writing SQL alone—experiments showed that context-passing costs outweighed the gains from specialization.

Third: after SQL runs, there's still dialect adaptation (syntax differences across databases), enum value matching (the user says "Beijing," but the database stores "Beijing City"), and reconciliation logic differences across teams. iDA's approach: let the model face unified ANSI SQL, and leave the remaining complexity for the engineering layer to digest.

Industry View

Supporters argue that both of iDA's contrarian decisions hold up. Using SQL over DSL acknowledges that model context is a scarce resource. Using single Agent over multi-Agent acknowledges that short pipelines are more reliable when task boundaries can't be cleanly carved up.

But the dissent deserves a hearing. This architecture is heavily dependent on ByteDance's self-built data middle platform DataWind, its internal data asset governance, and its vector retrieval systems. External companies lack this infrastructure—replication costs are prohibitive. "The engineering pipeline is the moat" is another way of saying "small companies, don't even think about copying the homework."

One more concern: iDA's public materials disclose no key metrics—accuracy, latency, nothing. The claim that "speed and accuracy come from convergence across the entire pipeline" reads more like engineering jargon that ordinary readers cannot independently verify.

Impact on Regular People

For enterprise IT: procurement standards for data analysis tools are about to change. You can no longer focus only on whether the model is GPT-4-level—you have to check whether the vendor has the unglamorous capabilities of dialect adaptation, field value recall, and business-logic reconciliation.

For individual careers: the barrier to data querying is falling for white-collar workers, but "asking the right question" is becoming more valuable than "knowing how to use the tool." The authority to define business logic will concentrate in the hands of those who truly understand the business.

For consumer markets: no near-term impact. This is an enterprise-facing tool with little connection to ordinary consumers' daily lives.