dd-parser-cleaner 0.8.0

A private, local LLM-powered data dictionary parser and entity mapper with automated cleaning.

dd-parser-cleaner

One-line summary
dd-parser-cleaner inspects incoming datasets, emits validated manifests describing structure and modalities, runs deterministic integrity checks, and writes a handshake file that downstream featurizers must read before transforming data.

Purpose

This package provides discovery and validation for enterprise datasets. It detects dataset type (cross-sectional, event-log, panel, homogeneous/bipartite/heterogeneous graph), tags attributes with roles and modalities, validates keys and joins, and produces actionable diagnostics and remediation hints. The canonical outputs are dataset manifest, attribute manifest, and handshake.json.

Quick start (workflow)

1. Discover package capabilities

from dd_parser_cleaner import get_package_info
info = get_package_info()

2. Run the parser (CLI or API) to produce:

• manifests/<dataset_id>.json (dataset manifest)
• attributes/<dataset_id>_attributes.json (attribute manifest)

3. Run the cleaner to validate manifests and produce:

• manifests/handshake.json

4. Featurizer must read manifests/handshake.json and proceed only if status == "ready".

Key capabilities

• Dataset discovery: auto-detects dataset_type and primary/time keys.
• Attribute tagging: emits role, time_dependency, granularity, modality, suggested_checks, generated_key_flag.
• Graph support: homogeneous, bipartite, heterogeneous graphs with entity/relationship maps.
• Longitudinal support: event-log vs panel; static vs dynamic attributes.
• Manifest emission: canonical JSON manifests for downstream deterministic featurization.
• Cleaner validations: monotonicity, lag consistency, cycle detection, relation consistency, URL/geo sanity checks.
• Handshake contract: handshake.json with status (ready | blocked | warnings).
• Config driven: behavior controlled by config.yaml flags.

Example artifacts

Example dataset manifest (snippet)

{
  "dataset_id": "orders_2026",
  "dataset_type": "event_log",
  "primary_key_spec": ["order_id"],
  "time_key_spec": "event_time",
  "entity_files": [],
  "relation_files": [],
  "panel_variable_map": null,
  "notes": "Order events from e-commerce pipeline",
  "validation_errors": []
}

Example attribute manifest entry

{
  "attribute_name": "order_id",
  "role": "subject_key",
  "time_dependency": "none",
  "granularity": null,
  "modality": "categorical",
  "suggested_checks": ["null_profile"],
  "generated_key_flag": false
}

Example handshake.json

{
  "status": "ready",
  "manifest_path": "manifests/orders_2026.json",
  "blocking_reasons": []
}

Where to find schemas and examples

• JSON Schema files (manifest validation): schemas/dataset_manifest.json, schemas/attribute_manifest.json, schemas/handshake.json
• Sample manifests and fixtures: tests/fixtures/manifests/ and tests/fixtures/csvs/
• Docs and design: USER_GUIDE.md, documents/, and docs/manifest.md

Important config flags (defaults)

Add or review these in config.yaml under a manifest section:

manifest:
  require_manifest_before_featurize: true
  use_case_questions_enabled: false
  graph_entity_limit: 5
  generate_surrogate_keys: true
  url_sample_size: 10

Handshake contract (featurizer requirements)

• Featurizer must read manifests/handshake.json before any transformation.
• If status == "blocked", the featurizer must refuse to proceed.
• If status == "warnings", the featurizer may proceed only after acknowledging and recording the warnings.

Migration and compatibility

• New manifest fields are additive and optional. Existing cross-sectional outputs remain unchanged during phased rollout.
• Recommended phased rollout:

1. Emit manifests and handshake while preserving legacy outputs.
2. Enable cleaner validators and handshake enforcement behind config flags.
3. Deprecate legacy outputs after one release cycle.

Troubleshooting (common validation failures)

• Missing primary key: parser will generate a surrogate key and set generated_key_flag; prefer providing explicit keys.
• Time key absent for longitudinal data: set time_key_spec or mark dataset as cross_sectional.
• Relation file join mismatch: ensure entity_key_spec matches keys referenced in relation files.
• Heterogeneous graph cycle detected: convert to acyclic tree or correct relationship files.
• Invalid URLs or geo addresses: check modality tags and sample rows flagged in diagnostics.

Each validation error includes severity, remediation, and sample_rows in the cleaner report.

How clients and agents should use get_package_info()

Use get_package_info() to discover:

• CLI commands and entry points
• manifest_schema_paths for validation
• handshake_spec and allowed status values
• supported_dataset_types and important config_flags

Treat get_package_info() as the canonical programmatic discovery endpoint.

Support and contribution

• Issue tracker: add issues at the repository issue tracker (link in get_package_info() output).
• Contributing: follow repository CONTRIBUTING.md for tests, fixtures, and schema updates.
• Contact: open an issue for integration questions or schema clarifications.

One-line blurb for top-level README

dd-parser-cleaner inspects datasets, emits validated manifests and a handshake file describing keys, time semantics, modalities, and graph structure, and provides deterministic diagnostics so downstream featurizers can safely and reproducibly transform data.

Existing quick links

• USER_GUIDE.md for usage details
• documents/ for methodology and internal design notes
• tests/notebooks/ for example notebook workflows