vasilisplavos/unboxgov-khmdhs

unboxgov — KHMDHS public-procurement dataset

Anomaly-detection over Greek public procurement, built from the KHMDHS OpenData feed. The dataset bundles the raw ingested records, cleaned analytical tables, detected patterns, generated reports, and the original contract PDFs in a single, atomically-versioned repository.

Indications, not accusations. The flags here surface patterns worth reviewing — never findings of wrongdoing. Always verify against the underlying records before drawing conclusions.

Repository layout

raw/
  <year>/*.parquet         # append-only raw KHMDHS records (see "Latest version")
  ingest_index.parquet     # per-day ingestion coverage ledger
metadata/
  contracts.parquet        # one row per contract
  contract_items.parquet   # one row per CPV section (join on contract_id)
  flags.parquet            # detected patterns (an indication may span contracts)
  orgs.parquet             # per-organisation (AFM) aggregates
pdfs/
  ...                      # original contract PDFs, sharded (see "PDFs")
pdf_index.parquet          # per-contract PDF lookup (status, sha256, shard, bytes)
site-assets/contracts.sqlite  # full original payload per contract (point-lookup DB)
site-assets/org_pages.sqlite  # precomputed per-organisation (AFM) page payloads
reports/
  <range>/...              # generated Greek-language report summaries
manifest.json              # repository state, file stats, paths
schema.json                # column-level data contract
README.md                  # this file

metadata/contracts joins metadata/contract_items on contract_id. CPV codes, quantities, and per-section pricing live on contract_items, not on contracts. See schema.json for column-level definitions and manifest.json for current row counts, byte sizes, and checksums.

Latest version of a contract (important)

The raw layer is append-only. A pipeline run never edits an existing parquet file — it adds a new one. When a contract's content changes upstream, a new row is appended with version = previous + 1; unchanged contracts are not rewritten. So a single reference_number can appear multiple times across raw/<year>/*.parquet, once per version.

To get the current state of each contract, keep only the highest version per reference_number (tie-break on the latest updated_at). Do not assume one row per contract.

Each raw row has: reference_number, submission_date, payload (the verbatim KHMDHS JSON envelope), content_hash, ingested_at, updated_at, version.

DuckDB:

SELECT * EXCLUDE (rn) FROM (
  SELECT *, row_number() OVER (
    PARTITION BY reference_number
    ORDER BY version DESC, updated_at DESC
  ) AS rn
  FROM read_parquet('raw/2026/*.parquet')
) WHERE rn = 1;

polars:

import polars as pl

latest = (
    pl.scan_parquet("raw/2026/*.parquet")
      .sort(["version", "updated_at"], descending=True)
      .unique(subset=["reference_number"], keep="first")
      .collect()
)

PDFs

Original contract PDFs live under pdfs/ (sharded into subdirectories), and are indexed by pdf_index.parquet (one row per contract: reference_number, status, bytes_raw, bytes_shrunk, sha256, shard, timestamps). PDFs are never embedded inside parquet and are never stored alongside the JSON payloads — binary blobs bloat columnar files and defeat the HTTP range-reads the parquet/SQLite layers depend on. Resolve a contract's PDF by looking up its shard in pdf_index.parquet.

Why everything lives in one repository

The analytical tables, the raw JSON records, and the contract PDFs are kept together rather than scattered across external storage so that every published snapshot is reproducible and versioned as a unit. Co-location means a given revision of the dataset always pairs the same raw inputs, derived tables, and source PDFs — there is no separate store that can silently drift out of sync.

Schema & manifest

schema.json is the column-level data contract; manifest.json records the current paths, row counts, byte sizes, and checksums for each artifact. Read those for machine-checkable details.