data-explorer
bởi google-gemini
Hồ sơ dữ liệu và khám phá đa năng. Sử dụng khi lần đầu tiếp xúc với bất kỳ bộ dữ liệu nào để hiểu cấu trúc, chất lượng và tiềm năng phân tích của nó.
npx skills add https://github.com/google-gemini/gemini-managed-agents-templates --skill data-explorerData explorer skill
Profile any tabular dataset (CSV, JSON, Parquet) and produce a structured summary the other skills can consume.
Workflow
- Scan workspace: list all data files in the workspace directory.
- Load and profile each file:
- Row count, column count
- Column names, data types, null counts, unique counts
- Basic statistics (min, max, mean, median, std for numerics)
- Value counts for categorical columns (top 10)
- Correlation matrix for numeric columns
- Assess data quality:
- Missing value percentage per column
- Potential data type issues (e.g., numbers stored as strings)
- Duplicate row detection
- Outlier detection (IQR method)
- Output a structured profile as JSON for downstream skills.
- Recommend analysis directions based on what you found.
Output format
{
"files": [
{
"filename": "customers.csv",
"rows": 91,
"columns": 7,
"schema": [
{"name": "CustomerID", "dtype": "object", "nulls": 0, "unique": 91},
{"name": "CompanyName", "dtype": "object", "nulls": 0, "unique": 91}
],
"quality": {
"missing_pct": {"Region": 0.60},
"duplicates": 0
},
"recommendations": [
"CustomerID is a unique string identifier",
"Region column has a high missing percentage (60%)",
"Can be joined with orders.csv on CustomerID to analyze customer behavior"
]
}
]
}
Key rules
- Never assume a specific dataset. Profile whatever is present.
- If no data files are found, inform the user and ask them to upload.
- Use
pandasfor profiling. It is pre-installed in the sandbox. - Use
select_dtypes(include=["object", "str"])for categorical columns. - For large files (>100K rows), profile a sample first and note the sampling.