silueta

Profile the shape of your data — never the values.

silueta turns a dataset — CSV, Parquet, and soon multi-tab Excel and databases — into a compact, machine-readable profile of its shape: types, nulls, cardinality, value-shape masks, semantic PII/PHI flags, and red-flag anomalies. The profile contains no raw cell values, by construction, so an AI agent (or a teammate, or a ticket) can work from it safely.

The motivating workflow: you get a vendor workbook full of PII/PHI and need to model it — dbt sources, Snowflake landing tables, relationships. Most of what modeling needs is how the data behaves, not what it says. silueta captures the behavior so an agent like Claude Code can scaffold the models for data it is never allowed to read.

$ uvx silueta scan patients.csv --report report.md
wrote profile.json
wrote report.md
  patients: 30 rows, 7 columns, 2 alerts

The profile tells the agent things like:

• mrn — VARCHAR, 100% unique, mask AAA-999999 → this is your key
• ssn — matches SSN structure at 100% with plausibility checks → sensitive, alert
• amount — typed as text but 100% castable numeric, magnitude 10^2 → DECIMAL, fix at landing
• clinic — constant column (and that is all it says about it)
• visit_date — dates spanning 2025 (year granularity only)

The contract

A silueta profile never contains:

• raw cell values, sample rows, or frequent-value lists
• min/max for strings or dates (those are raw values) — numerics report order-of-magnitude and precision facts, dates report year spans
• serialized sketches or value hashes (hashing is not anonymization)
• any statistic derived from values observed in fewer than k rows (default 5); constant columns report nothing but their constancy

This contract is enforced by tests that scan every output for source values, and it travels inside the artifact itself (profile.json → silueta.contract).

What it is and is not: silueta protects against accidental context leakage — values ending up in an LLM conversation, a log, a pasted ticket. It does not claim formal anonymity (differential privacy) against an adversary reconstructing small populations from aggregates. The wording is "no raw values by construction," never "anonymized."

Status

Early alpha — CSV, Parquet, and multi-sheet Excel profiling work today, including messy-workbook recovery: title/footer rows skipped, multi-row headers merged, numbers-stored-as-text promoted to real types with confidence scores, non-tabular sheets flagged and skipped. On deck (see issues): cross-table foreign-key candidates, dbt/Snowflake scaffold emitters, database connections.

Install

uvx silueta            # zero-install run
uv tool install silueta
pipx install silueta

Agent setup

silueta is built for agent-driven data work: the profile is the interface an agent reads, and the repo ships a Claude Code skill plus a lockdown recipe that blocks raw data reads (Read-tool denies + a shell hook) while allowlisting the profiler — the agent models data it is never allowed to read.

License

MIT