Reviewing the output

The package returns a ranked JSON array of issues and writes nothing to a catalog. The review of those issues happens in one of two places.

Inside the Metadata Editor

The reviewer ships as an optional add-on to the Metadata Editor's FastAPI service. When it is enabled, Assess metadata on a project's Issues page runs the pipeline and lists the AI-generated issues with their suggested corrections, and each one can be commented on, accepted, rejected, dismissed, or applied through the Editor's issues workflow. The Editor's installation guide describes the setup, and the assess metadata page describes the review.

Outside the Metadata Editor

When the package runs on its own, the results of a catalogue review can be assembled into one table for triage in a spreadsheet or in the organization's own review tool. The table has one row per record, with the issue array stored as a JSON string:

import json
import pandas as pd
from ai4data.metadata.reviewer import MetadataReviewerClient

client = MetadataReviewerClient.from_openai(model="gpt-4o", api_key="sk-...")

rows = []
for project_name, metadata in catalogue.items():  # a dict of records
    job = client.submit(metadata)
    issues = job.wait_sync(timeout=300)
    rows.append({
        "ME_project": project_name,
        "detected_issues": json.dumps(issues, ensure_ascii=False),
    })

pd.DataFrame(rows).to_excel("review_input.xlsx", index=False)

Each object inside the JSON array carries detected_issue, issue_category, issue_severity, current_metadata, and suggested_metadata, the schema the pipeline produces. Sorting by issue_severity gives the triage order, and the key paths in current_metadata and suggested_metadata identify the field to change. Applying an accepted suggestion to the metadata source is a separate, deliberate step; the metadata curation cookbook gives a review procedure and a decision log for it.

No bundled review interface

Earlier versions of this manual described a standalone HTML review board. That file is not distributed with this repository.