Skip to main content

Overview

rulesgen is a secure rule-processing service for synthetic data workflows. It accepts rule input as either natural_language or a restricted DSL, translates natural_language requests into an untrusted semantic_frame plus DSL candidate, validates the DSL into a compiled_rule, supports local execution_preview, and can execute full dataset generation as a tracked job.

The service is designed around a clear trust boundary: LLM output is never trusted directly. Natural-language output and DSL candidates remain untrusted until parser and validator checks succeed, and diagnostics are part of the contract at every stage.

Who this documentation is for​

  • Teams authoring rules for synthetic data workflows.
  • Operators evaluating local preview and full dataset-generation paths.
  • Application teams embedding the Python library or HTTP API.
  • Platform teams configuring LLM, guardrail, storage, and execution backends.

What rulesgen does​

The core workflow is staged so each step can be inspected independently:

  1. Parse rule input into a semantic_frame, DSL candidate, and diagnostics.
  2. Treat natural_language output and every DSL candidate as untrusted until validation succeeds.
  3. Compile validated DSL into a compiled_rule artifact.
  4. Run an execution_preview against a sample row and seed.
  5. Generate a target dataset as a tracked job.
  6. Inspect diagnostics, generated artifacts, and job metadata.

rulesgen can be used through the HTTP API or directly as a Python library. The HTTP service provides endpoints for parsing, compilation, preview, dataset uploads, dataset generation, job polling, and artifact download. The library API exposes the same core capabilities for in-process use.

Safety model​

rulesgen is not a direct natural-language-to-Python executor. A natural_language rule is translated into an LLM-produced semantic_frame and DSL candidate, and both remain untrusted until validation succeeds. The compiler accepts only a restricted Python-expression subset and a runtime helper whitelist. The preview executor is intended for fast local feedback; full dataset generation runs through either the subprocess dataset executor or the optional Alibaba OpenSandbox adapter.

Pre-LLM guardrails scan natural-language input for prompt injection and jailbreak attempts. Blocked requests return a standard Problem Details response without exposing scanner internals to the caller.

Start here​