open-data-quality

name: open-data-quality description: "Comprehensive open data quality validator for two audiences: data analysts who need to assess whether a dataset is ready to use, and public administrations who want to self-evaluate their published data. Automatically adapts based on input type: (A) local CSV file only — performs file-level structural and content checks; (B) CKAN/open data portal dataset — adds metadata completeness, resource accessibility, URL reachability, and DCAT-AP compliance (supports all national profiles: DCAT-AP 2.x baseline, IT, BE, NL, DE, FR, UK, ES, and others). Always use this skill when the user mentions: data quality, validate dataset, check CSV, open data compliance, metadata audit, CKAN dataset review, "is this data usable?", or whenever a CSV file or CKAN dataset ID/URL is provided for quality assessment. Produces severity-ranked reports (blocker / major / minor) with concrete fixes, quality score, and a plain-language summary for non-technical stakeholders."

Open Data Quality Validator

Overview

This skill validates open data quality for two contexts and two audiences:

Input contexts (auto-detected):

• Local CSV — file on disk, no portal metadata available
• CKAN / open data portal dataset — dataset ID or URL

Target audiences:

• Data analysts — "Can I trust and use this data?"
• Public administrations — "Does our publication meet open data standards?"

Input Detection

Before running any checks, identify what you have:

LOCAL CSV only → run odq-csv (phases 0–4)
CKAN/portal   → run odq-ckan (phases 0–6, optionally --download for CSV too)

Signs of a CKAN dataset: user provides a dataset ID, a CKAN URL (/dataset/...), or asks to retrieve a dataset from a portal. If uncertain, ask: "Do you have the file locally, or should I retrieve it from an open data portal?"

Installation

The tool runs on any OS (Windows, macOS, Linux) via uvx — no permanent install needed. If uv is not available, install it first:

# Install uv (once)
curl -LsSf https://astral.sh/uv/install.sh | sh        # macOS / Linux
winget install --id=astral-sh.uv -e                    # Windows

Codex / sandboxed environments: if uvx fails with Invalid cross-device link (EXDEV), the sandbox filesystem isolates temp and cache directories across different mounts. Run outside the sandbox, or set:

UV_LINK_MODE=copy UV_NO_CACHE=1 uvx --from git+https://github.com/ondata/open-data-quality odq-ckan ...

Windows note: set PYTHONUTF8=1 to avoid encoding errors on cp1252 terminals:

# PowerShell
$env:PYTHONUTF8=1; uvx --from git+https://github.com/ondata/open-data-quality odq-csv data.csv

# Command Prompt
set PYTHONUTF8=1 && uvx --from git+https://github.com/ondata/open-data-quality odq-csv data.csv

Usage

Local CSV (phases 0–4)

# Run directly — no install needed
uvx --from git+https://github.com/ondata/open-data-quality odq-csv data.csv

# With JSON and Markdown output
uvx --from git+https://github.com/ondata/open-data-quality odq-csv data.csv --output-json report.json --output-md report.md

CKAN portal dataset (phases 0–6)

# Metadata + accessibility only (fast)
uvx --from git+https://github.com/ondata/open-data-quality odq-ckan PORTAL_URL DATASET_ID

# Metadata + accessibility + CSV validation
uvx --from git+https://github.com/ondata/open-data-quality odq-ckan PORTAL_URL DATASET_ID --download

# With output files
uvx --from git+https://github.com/ondata/open-data-quality odq-ckan PORTAL_URL DATASET_ID --download --output-json report.json --output-md report.md

Validation Modes

Default to Standard Check unless the user asks otherwise.

What Gets Checked

For a full breakdown of all checks per phase, see scripts/README.md.

Local CSV (phases 0–4): file validity, encoding, BOM, column names, wide format, aggregate rows, decimal separators, date formats, units in cells, placeholders, reference codes.

CKAN portal (phases 0–6): all CSV checks (with --download) plus DCAT-AP profile detection, metadata completeness, per-resource format/license/URL, HTTP accessibility, declared vs actual encoding, update frequency consistency.

Qualitative assessment (LLM, after scripts, requires data content): title discoverability, title ↔ description coherence, description ↔ content coherence, content ↔ declared update frequency, dataset usefulness, unit of observation, column name semantics, provenance and methodology, joinability.

Qualitative Assessment (LLM-only, after scripts)

After running the scripts, perform a qualitative assessment using the data content and metadata already read. This step applies only when data content is available (local CSV, or CKAN with --download). It is done by the LLM — not by a script — and produces observations, not scores.

Run these five checks in order:

1. Title discoverability Read the dataset title and the column names + a sample of values. Ask: would someone searching for this data find it with the title as written? Does the title use clear, domain-relevant keywords, or is it vague/internal jargon (e.g. "Export_2024_v2_final")?

2. Title ↔ description coherence Read title and description together. Do they describe the same subject? Flag if the description contradicts the title, is a copy-paste of it, or adds no meaningful context.

3. Description ↔ content coherence Compare the description text with the actual column names and sample values. Does the description accurately reflect what the data contains? Flag mismatches (e.g. description says "hospital admissions" but columns are about road accidents).

4. Content ↔ declared update frequency If update frequency is declared in metadata, look at date columns in the data. Does the latest date in the data match what you would expect for the declared cadence (daily, monthly, annual…)? Flag if the data looks stale or if the date range is inconsistent with the declared frequency.

5. Dataset usefulness Consider: does the combination of row count, descriptive columns, and update cadence make this dataset genuinely useful? Flag if the dataset is too sparse (too few rows or columns for the declared scope), if descriptive fields are mostly codes with no labels, or if the data sample suggests the published extract is incomplete.

6. Unit of observation From column names and sample values, can you immediately tell what one row represents (a municipality? a transaction? a person? a time period?)? Flag if the unit is ambiguous or mixed (e.g. some rows are totals, others are details), as this is one of the most common and costly usability problems in open data.

7. Column name semantics Are column names self-explanatory to an external user, or are they internal acronyms/codes (e.g. FL_ATT, IMP_NET_PRG, cod_val_2)? Flag columns whose meaning cannot be inferred without a data dictionary. This is particularly common in Italian PA data and is a real barrier to reuse.

8. Provenance and methodology From the metadata and description, is it clear how the data was collected or produced (administrative records, survey, sensor, aggregation…)? Without provenance, data cannot be trusted even if technically well-formed. Flag if the origin is absent or vague.

9. Joinability Does the dataset contain reference codes (NUTS, ISTAT municipality codes, CIG, CUP, fiscal codes, ISO country codes…) that allow joining it to other datasets? Note which codes are present and whether they appear well-formed. Joinability is a core value proposition of open data as an ecosystem.

How to report: add a ## Qualitative Assessment section to the report (see template below). For each check, write one sentence: what you observed and whether it raises a concern. Use Good / Acceptable / Poor — not blocker/major/minor. If data content is not available (metadata-only run), skip this section entirely.

Report Template

# Open Data Quality Report
**Dataset**: [title or filename]
**Source**: [Local CSV | Portal: URL]
**DCAT-AP profile**: [baseline 2.x | IT | DE | FR | BE | NL | UK | ES | unknown]
**Validated**: [YYYY-MM-DD]
**Mode**: [Quick Triage | Standard Check | Full Audit]

## Summary
[2–3 plain-language sentences: who published this, is it usable, main problems.]

## Quality Score: XX/100

## Blocker Issues ⛔
## Major Issues ⚠️
## Minor Issues ℹ️

## Score Breakdown
| Dimension | Score | Notes |
|-----------|-------|-------|
| Accessibility | /20 | Portal only |
| Metadata completeness | /20 | Portal only |
| File format compliance | /15 | |
| Data structure quality | /20 | |
| Data content quality | /25 | |

## Qualitative Assessment
*(skip if data content not available)*
| Check | Observation | Rating |
|-------|-------------|--------|
| Title discoverability | | Good / Acceptable / Poor |
| Title ↔ description | | Good / Acceptable / Poor |
| Description ↔ content | | Good / Acceptable / Poor |
| Content ↔ update frequency | | Good / Acceptable / Poor |
| Dataset usefulness | | Good / Acceptable / Poor |
| Unit of observation | | Good / Acceptable / Poor |
| Column name semantics | | Good / Acceptable / Poor |
| Provenance and methodology | | Good / Acceptable / Poor |
| Joinability | | Good / Acceptable / Poor |

## Recommended Actions
[Numbered, by severity, with code examples]

## What This Means for You
**As a data analyst**: [Can I use this? What to fix before analysis?]
**As a publisher**: [What must be fixed before republishing?]

Severity Levels

Exit Codes

References

• DCAT-AP 2.x
• RFC 4180 — CSV · ISO 8601
• Tidy Data · Frictionless Table Schema
• IT: DCAT-AP_IT · AgID
• DE: DCAT-AP.de · FR: schema.data.gouv.fr
• BE: DCAT-AP_BE · NL: DCAT-AP_DONL
• UK: data.gov.uk profile · ES: datos.gob.es

Extended References

• references/dcat-ap-profiles.md — Field-by-field comparison across all national profiles
• references/common-antipatterns.md — Universal + locale-specific patterns with fixes
• references/dcat-ap-it-fields.md — Detailed DCAT-AP_IT field list (Italian PA)
• scripts/README.md — Full tool documentation, all options and examples