datasentry

中文导读：DataSentry 会先自动发现数据质量问题，再给出样本、比例、置信度等证据。修复采用 propose → preview → apply to a copy → verify → rollback 的保守流程。检测与评分不依赖 LLM，AI 只作为可选辅助，数据可以完全留在本机。

Try it in 30 seconds

Install the current PyPI release and run the zero-config product tour:

pip install --upgrade datasentry-ai
datasentry demo

datasentry-demo is also available as a direct console alias.

The demo generates synthetic dirty data, runs the built-in detectors, exports JSON + HTML reports, applies one safe repair to a copy, re-scans the repaired copy, and prints a rollback command. It needs no dataset, cloud service, API key, or LLM.

synthetic dirty CSV
        ↓
39 deterministic detectors
        ↓
evidence-backed issues + quality score
        ↓
preview → repaired copy
        ↓
re-scan → verify new/persistent issues

Prefer scanning your own data immediately?

curl -L https://raw.githubusercontent.com/Jackxiaozhiren/datasentry/main/demo-data/orders.csv -o orders.csv
datasentry scan orders.csv
datasentry issues list --severity high

Or launch the interactive interfaces:

datasentry          # terminal UI
datasentry-server   # Web UI + REST API at http://localhost:8000/ui/

Related MCP server: aegis-dq

Start with your problem

The default path is intentionally low-commitment: scan first, inspect evidence, then decide whether a finding deserves a permanent rule or a repair.

Why DataSentry exists

Most data-quality tools are excellent once you already know the expectations, checks, or contracts you want to enforce. Real incidents often start one step earlier: you do not yet know what is wrong.

DataSentry is built around the complete remediation loop:

Find → Explain → Fix safely → Verify

• Find — discover common quality problems without writing every rule first.

• Explain — attach samples, affected counts/ratios, detector evidence, and confidence.

• Fix safely — preview changes and apply repairs to a copy instead of mutating the source.

• Verify — re-scan the repaired copy and surface persistent or newly introduced issues.

DataSentry fits best when you are looking at data you do not fully trust yet, want evidence before codifying checks, or need a conservative remediation path. It complements expectation- and contract-driven tools rather than requiring you to replace them.

What it catches automatically

DataSentry ships with 39 deterministic detectors covering common failure modes such as:

• missing and placeholder values;

• invalid emails, URLs, dates, and encodings;

• duplicate identifiers and uniqueness violations;

• inconsistent categories and cross-field contradictions;

• foreign-key and referential-integrity problems;

• numeric and statistical outliers;

• schema, row-count, score, and issue-distribution drift.

Every scan produces an evidence-backed issue list and a six-dimension quality score across completeness, validity, uniqueness, consistency, integrity, and timeliness.

Safe repair, not blind mutation

# inspect the highest-severity findings
datasentry issues list --severity high

# propose a repair without changing data
datasentry repair propose <issue_id> --file orders.csv

# preview the exact effect
datasentry repair preview <issue_id> --file orders.csv

# apply to a repaired copy; the original is not overwritten
datasentry repair apply <issue_id> --file orders.csv

# re-scan the repaired copy and detect regressions
datasentry repair verify <run_id>

# inspect or undo the repair
datasentry repair diff <run_id>
datasentry repair rollback <run_id>

Repairs are fingerprinted, auditable, and reversible. AI-generated repair proposals remain human-approved state changes.

Quality gates for CI

datasentry scan orders.csv --fail-on high

For GitHub repositories, reuse DataSentry's maintained workflow_call gate instead of duplicating installation and exit-code handling:

jobs:
  datasentry:
    uses: Jackxiaozhiren/datasentry/.github/workflows/datasentry-quality-gate.yml@main
    with:
      path: data/orders.csv
      fail_on: high

See docs/GITHUB_ACTIONS.md for inputs, artifacts, security boundaries, and version-pinning guidance.

Reports can be exported as JSON, Markdown, HTML, JUnit, and SARIF. The GitHub Actions example fails the workflow on severe findings while still uploading an HTML report for review.

Give AI agents deterministic data-quality tools

DataSentry includes an MCP stdio server:

datasentry mcp --project /path/to/project

MCP-capable clients can scan files, inspect evidence-backed issues, read quality scores and trends, compare drift, validate contracts, manage scheduled jobs, and call DataSentry tools without bypassing the same underlying safety rules used by the CLI and REST API.

Copy-paste setup recipes for VS Code and Claude Desktop are in docs/MCP.md.

Boundary: AI may propose; humans approve state-changing repairs.

Where it fits next to popular data-quality projects

This is a positioning guide, not a winner/loser feature scorecard. These projects solve overlapping but different jobs; check their upstream documentation for current capabilities.

DataSentry is intentionally not trying to replace a metadata catalog, lineage platform, or every validator. Its focus is narrower: find bad data, show why it was flagged, and close the repair loop without gambling on the source.

Local-first by design

• deterministic detection and scoring run locally;

• DuckDB powers core local execution;

• OpenAI/Ollama assistance is optional;

• PII redaction, encrypted mappings, and LLM audit records are available when AI is enabled;

• the original source file is not overwritten by repair workflows.

Data sources

• CSV, Parquet, JSONL, XLSX

• DuckDB and SQLite

• PostgreSQL and MySQL

• s3://, gs://, and az:// objects

• single files, batches, and globs

History and drift

Persist scans and compare data over time:

datasentry drift latest orders
datasentry score

Tracked signals include schema changes, row-count movement, quality-score changes, and issue-distribution drift.

Architecture

flowchart LR
    Sources[Files / DBs / cloud objects] --> DuckDB[Local execution]
    DuckDB --> Detect[39 detectors]
    Detect --> Evidence[Evidence fusion]
    Evidence --> Score[6-dimension score]
    Score --> Reports[Reports / history / gates]
    Reports --> CLI[CLI / TUI]
    Reports --> Web[Web / REST]
    Reports --> MCP[MCP]
    Evidence --> Proposal[Repair proposal]
    Proposal --> Preview[Preview]
    Preview --> Apply[Apply to copy]
    Apply --> Verify[Verify by re-scan]
    Verify --> Rollback[Rollback artifact]
    LLM[Optional OpenAI / Ollama] -. proposes .-> Proposal

Reproducible benchmark

uv sync
uv run python benchmarks/bench_scan.py 1000000 42

The benchmark generates synthetic dirty data and measures profiling, detection/fusion/scoring, numeric-outlier detection, JSONL reading, sampling, score drift, and memory high-water marks. See docs/BENCHMARKS.md.

Development

uv sync
make check          # lint + mypy --strict + tests/coverage
make demo           # exercise the public datasentry demo path
make bench          # benchmark
make build          # distributions

Contributing

Useful contributions include new detectors, connectors, integration examples, reproducible benchmark cases, documentation/translations, minimal bug reproductions, and CLI/TUI/Web usability improvements.

Want a small first contribution? Start with good first issue #14 or good first issue #15.

See CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, and ROADMAP.md. Small contributions are tracked with good first issue and help wanted labels.

Documentation

• Project site and live report

• docs/FAQ.md — evaluation, safety boundaries, and how DataSentry fits alongside other data-quality tools

• examples/ — scenario-first runnable examples

• docs/MCP.md — VS Code and Claude Desktop MCP setup

• docs/GITHUB_ACTIONS.md — reusable GitHub quality gate

• examples/integrations/github-actions/ — copy-paste CI gate

• docs/BENCHMARKS.md — benchmark policy

• docs/DEVELOPMENT.md — engineering notes

• Detect → fix → verify / 中文版

License

Apache-2.0 — see LICENSE.