DataWorks AI Data Generation: One JSON Config Yields Simulation-Grade Business Data

Traditional Data Generation vs. AI Data Generation

Creating realistic business data for testing, demos, algorithm training, and compliance has always been expensive: manual entry looks fake, external scraping is risky, and script-based generation is hard to maintain. For complex wide tables, traditional methods hit three bottlenecks: cross-field mismatches, inability to guarantee consistency constraints (e.g., detail sums, multi-party balances), and scripts that cannot become reusable team assets.

DataWorks Data Integration AI Data Generation (released with the 2026-10-08 version) turns data creation into a pipeline within data integration. A single JSON configuration suffices: the LLM orchestrates the business "plot" while the rule engine guarantees "physical" self-consistency, producing simulation-grade business wide tables in minutes that pass business validation and are ready for downstream consumption.

Key Advantages

Minute-level delivery : One JSON config creates a task; once configured it runs in minutes, supports iterative changes. Traditional development and debugging take days.

Behavior feels real : The LLM infers complete "plots" covering behavior style, environment, events, and temporal transitions — no longer obviously fabricated numbers.

Cross-field self-consistency : The rule engine enforces declared constraints — aggregation, balance, monotonicity, mutual exclusion — all self-consistent, passing business validation directly.

Temporal coherence : State machine plus previous-moment context ensures each entity's behavior links sequentially and is replayable.

Elastic scale : From 1 to thousands of entities differs by only a few config lines; scaling requires no code changes.

Ready-to-use : Lands directly into MaxCompute partitioned wide tables (partitioned by business date), eliminating CSVs and scattered temporary tables.

Universal Capability Model: Three-Layer Abstraction, Industry-Agnostic

Why can AI-generated data both "pass for real" and "fit perfectly"? The core is a three-layer abstraction model :

Field Dictionary (Schema Layer) : Defines the wide table structure — column names, types, groupings, value domains. It is the single schema source.

Business Semantic Profile (Plot Layer) : The LLM outputs only a few "decision" parameters answering "what is this entity experiencing at this moment" — behavior style, lifecycle stage, environmental conditions, events and anomalies, individual attributes, operation modes.

Rule Expansion (Physics Layer) : The rule engine deterministically expands the profile into all signal columns per user-declared physical/business rules — aggregation consistency, directional consistency, value domains, monotonic constraints, mutual exclusion, balance constraints.

Switching industries = swapping a field dictionary + a rule set. The six-step pipeline, concurrency and partitioning mechanisms, and scheduling/operations are fully reused — no re-development needed.

Industry Scenario Matrix: Same Engine, Swap Dictionary for Reuse

Any data in the form of "entity × time-series × coupled fields" wide tables can plug into AI Data Generation. The built-in IoV template illustrates the three-layer model:

Field Dictionary : 800+1 columns covering 7 groups — BMS, voltage sampling, electric drive, thermal management, etc.

Semantic Profile : LLM outputs only 8 decision parameters (driving style, 7-phase trip stages, road conditions, fault injection, etc.).

Rule Expansion : Executes 14 physics consistency rules (e.g., pack voltage ≈ sum of cell voltages, SOC change direction matches current, three-phase current sum ≈ 0).

One config yields "1 vehicle × 200 continuous time steps" or "100 vehicles × 100 steps = 10,000 rows" of simulation datasets.

Other applicable scenarios include:

Embodied Intelligence / Behavioral Data : Multi-modal sensor and action time-series wide tables; task intent, action primitive sequences, environment interaction events, force feedback; joint limits, end-effector trajectory smoothness, force-position coupling consistency, timestamp monotonicity.

Financial Risk Control / Anti-Fraud : Transaction and account behavior streams; spending style, lifecycle, attack scripts; lending balance, limit mutual exclusion, temporal monotonicity.

Medical / Vital Signs : Monitor time-series wide tables; activity state, medication events, individual baselines; physiological value domains, indicator linkage, sampling monotonicity.

Industrial IoT / Equipment : Production line sensor time-series wide tables; operating phases, load, degradation events; energy conservation, three-phase balance, counter monotonicity.

Retail / User Behavior : Tracking and order behavior wide tables; crowd profiles, promotional events, channels; inventory conservation, amount = unit price × quantity.

How It Works: LLM Handles "Plot", Rule Engine Handles "Physics"

The LLM answers only a few "decision" parameters (what this entity is experiencing at this moment); all remaining signal columns are deterministically derived by the rule engine per your declared rules.

Why not let the LLM do the math? Deterministic calculations across hundreds of columns — sums, balances, monotonicity — suffer from uncontrolled arithmetic errors in LLMs. The rule engine uses numerical computation to guarantee hard constraints stably — making every column verifiable and reproducible .

Generation Pipeline: Six Steps in One Task

Seed generation, profile inference, rule expansion, and wide-table write happen in a single closed-loop task — no source tables required :

Data Generator : Four column generation modes produce seed records.

KeyBy Partitioning : Same entity always routes to the same concurrency slot, preserving time-series order.

Pre-processing : Initializes each entity's state and previous-moment context.

LLM Profiling : Calls the Tongyi model to output business semantic profile (strict JSON).

Rule Expansion : Deterministically expands into all signal columns per your declared rules.

Wide Table Write : Lands in MaxCompute, partitioned by business date.

The platform automatically injects resource groups, CUs, and LLM managed authentication — you never touch any secrets .

Elastic Scale and Cost Control

Two modes, choose as needed:

Single-entity deep time-series : Profile inferred step-by-step, state linked step-by-step, most coherent temporally. Suits algorithm tuning, fine-grained anomaly scenario construction.

Thousand-entity concurrent volume (recommended) : Entities run in parallel, each entity serial. Suits population-level simulation datasets, dashboard statistical analysis.

Quality and cost dual optimization : State machine ensures same entity behavior doesn't jump, replayable; LLM token consumption grows linearly with record count, supports batched generation, partitioned overwrite writes, transparent cost, re-runs produce no duplicate data.

Note: This capability focuses on "time-series coupled wide table" simulation. Pure static dimension tables or simple random data without temporal coupling are not target scenarios. Profile quality relates to prompt design; recommend small-scale trial run first.

Get Started Now

No data-generation code to write; single-entity fine-grained inference or thousand-entity concurrent volume, available now:

Write Config : Describe entity scale and business rules in one JSON config (existing industry templates can be applied directly).

Create Task : Create an AI Data Generation task (platform auto-injects resource group and LLM auth, no secrets needed).

Use Data : Minutes later, pull data from MaxCompute partitioned wide tables; downstream consumes immediately.