QiyanAgent 企小喵实证研究智能体 Download Client
Data Cleaning Agent

Diagnose the data first, then let the Data Cleaning Agent generate rules.

The Data Cleaning Agent diagnoses data first, then applies traditional, machine-learning, or LLM-based cleaning rules. It records the rules, anomaly handling, and quality checks; cleaned outputs are delivered as new files or tables and do not overwrite the original data.

Applies to QiyanAgent V3.5. Last updated: September 14, 2026.

Data Cleaning Agent steps: preserve the raw data, review the cleaning package and built-in Skills, enter cleaning goals and acceptance criteria, and send the task.

This is a QiyanAgent V3.5 interface screenshot. The callouts mark capability selection, Skills configuration, task entry, and task submission.

Before You Start

  • Open a project and place the data to clean in the project folder, or import it into the current DuckDB.
  • Keep the original data and do not overwrite source files.
  • State key fields, primary keys, time fields, units, variable definitions, and fields that must not be changed.
  • If cleaning standards already exist, put the rule document or data dictionary in the project folder first.

Steps

  1. In the Data Cleaning workspace, review packages for regex cleaning, missing values, duplicate records, type conversion, text normalization, category mapping, feature construction, entity alignment, fuzzy deduplication, anomaly detection, text classification, model imputation, structured extraction, and privacy redaction.
  2. Activate a cleaning package manually, or describe the problem and let the Agent decide which capability to load.
  3. To review capabilities, open Settings → Skills and switch to the Data Cleaning Agent. Built-in Skills sync automatically; use Import only to add cleaning methods.
  4. Enter a cleaning prompt with data location, field definitions, cleaning goals, output format, and quality checks.
  5. Click the send button and let the Data Cleaning Agent diagnose first, then generate rules and execute cleaning.

Acceptance Checks

The output should include before-and-after comparisons, rule descriptions, anomaly-handling rationale, row-count changes, field changes, and a quality report. Cleaned outputs should be saved as new files or tables, not overwrite the original data.