diavgis/greek-public-procurement

Diavgis — Greek Public Procurement, cross-linked (2026-03-01 → 2026-08-16)

Greek public procurement is published across three separate official systems that do not talk to each other: Διαύγεια (every administrative act, keyed by ΑΔΑ), ΚΗΜΔΗΣ (the central procurement registry, keyed by ΑΔΑΜ) and EU TED. A tender found in one cannot be followed to its award, its contract or its final payment in another. This dataset is those three sources harvested, normalized and cross-linked into reconstructed contract lifecycles.

⏱ This is a periodic snapshot, not a live feed

Snapshot taken 2026-08-16, covering acts issued 2026-03-01 → 2026-08-16. The corpus grows by roughly 1.300–1.700 acts every day, and the enrichment crons keep filling in amounts, CPV codes and supplier ΑΦΜs on records that are already in this file. So a row here can be both correct and out of date.

For the latest data, go to the source — do not re-crawl this dataset or the site:

you want	use
today's acts, current amounts, live lifecycle chains	https://diavgis.gr — every act, entity and chain has a permanent page
programmatic access, always current	https://diavgis.gr/docs — REST API + MCP server, free tier, one key
agent-readable text instead of HTML	send Accept: text/markdown to any page, or read https://diavgis.gr/llms.txt
a refreshed bulk file	this dataset, re-published periodically — check Last modified above

This snapshot is here so you can start without hitting an API 700.000 times. Once you need current data, the site and the API are the authoritative surface and are cheaper for everyone.

file	rows	size	what it is
acts.parquet	923,503	56 MB	one row per procurement act
entities.parquet	59,290	1.4 MB	legal entities (buyers and suppliers) resolved by ΑΦΜ
chains.parquet	438,693	3.3 MB	membership of acts in reconstructed lifecycle chains
act_cpv.parquet	806,392	6.1 MB	every CPV code attached to an act, with its label
act_suppliers.parquet	551,252	4.9 MB	co-awardees (framework agreements, batch payments)

2,779,130 rows in 72 MB — the same corpus was 474 MB as CSV.

By source: ΚΗΜΔΗΣ 467,879 · Διαύγεια 450,628 · TED 4,996. By lifecycle stage: award 282,589 · payment 264,845 · contract 162,768 · tender 135,353 · request 77,948. 862,627 acts carry a monetary amount.

Privacy — read this before using acts.parquet

Greek transparency law publishes payments to natural persons alongside payments to companies. This release does not reproduce that.

• 78,323 acts are marked redacted = 1. Their counterparty is a natural person. For every one of those rows supplier_afm is empty — the tax ID of a private individual is not in this file.
• entities.parquet contains only ΑΦΜs that appear on at least one non-redacted act. 2,001 entities reachable only through redacted acts were removed, because they cannot be distinguished from natural persons.
• act_suppliers.parquet excludes every row belonging to a redacted act.

Seven ΑΦΜs referenced in acts.parquet have no row in entities.parquet. That is deliberate: the entity filter above is conservative, and it is the right direction to err in for a permanent release. Join with a left join and expect a handful of misses.

The redacted acts themselves are kept — the amount, the buyer, the date and the CPV are legitimate public spending facts. Only the identity of the private individual is withheld. Please do not attempt to re-identify redacted counterparties from other sources.

Columns

acts.parquet carries the fields the public Diavgis API already publishes, plus spend_repr — which only makes sense in a bulk release, and which you need in order to sum correctly (see below). Nothing else was added: no storage keys, no content hashes, no normalization internals.

Changed in this release (2026-08-18). Five procurement-signal columns — procedure_type, bids_count, award_criteria, contract_type and legal_context — were present up to the 2026-08-16 release and are not in this one. They are served per record through the API and on the site; see Live access below.

The procurement method is still open data. The OCDS 1.1 feed publishes tender.procurementMethod — the standard's own codelist, open / selective / limited — for every release, free and without a key: https://api.diavgis.gr/ocds. What is no longer in this bulk file is the verbatim Greek wording of the procedure and the bid/criteria detail.

Also absent since 2026-08-14: contract_start, contract_end, contract_duration (present up to the 2026-08-06 CSV release).

Notable columns:

• id — primary key. Διαύγεια acts use the ΑΔΑ; ΚΗΜΔΗΣ acts use kimdis:<ΑΔΑΜ>; TED uses ted:<id>.
• ada, adam, esidis_id — the identifiers in each official system, where present.
• kind — request | tender | award | contract | payment.
• buyer_afm, supplier_afm — 9-digit Greek tax IDs, joinable to entities.parquet.
• amount_net, amount_gross, currency, contract_budget_net. Read the VAT note below before summing either one.
• cpv_primary — main CPV code; act_cpv.parquet holds the full set per act.
• nuts — NUTS region code. region_inferred = 1 means it was derived, not published at source.
• eu_funded — EU co-financing flag, where the source declares it.
• The other ΚΗΜΔΗΣ procurement signals (bids_count, procedure_type, award_criteria, contract_type, legal_context) are not in this file — see the note above. For the procurement method as an OCDS codelist, use the OCDS feed; for the full detail, the API or the site.
• quality_flags, reconcile_flags, amount_check — per-record validation trail.
• cancelled, revoked, superseded, prev_ref — lifecycle state at the source.
• spend_repr — 1 when the amount represents actual spend. Use this before summing. TED carries estimated multi-year framework budgets; summing them as spend inflates totals badly.
• redacted — 1 when the counterparty is a natural person (see Privacy).

chains.parquet has chain_id, act_id, role, confidence (0–1) and method — how the link was made. Filter on confidence if you need only high-certainty links.

⚠ VAT: amount_net is not always net — read this before you sum anything

This is the single most important caveat in the release, and it was measured, not assumed.

ΚΗΜΔΗΣ publishes both figures and this dataset carries them faithfully: amount_net is net of VAT and amount_gross includes it (237k of 238k ΚΗΜΔΗΣ award/contract rows carry both).

Διαύγεια publishes ONE number and does not say which it is. The source field (awardAmount.amount on an award, sponsor[].expenseAmount.amount on a payment) has no net/gross distinction and no VAT rate. It lands in amount_net because that is the only amount column the source fills — 299,654 Διαύγεια rows carry amount_net with amount_gross empty.

Measured over 2,732 chains that hold both a ΚΗΜΔΗΣ act (with distinct net and gross) and a Διαύγεια act, the ratio of the Διαύγεια amount to the ΚΗΜΔΗΣ net is the Greek VAT table almost exactly:

ratio	share
1.24	72.0%	VAT 24% (standard) — the Διαύγεια figure is GROSS
1.00	17.9%	published net
1.13	3.7%	VAT 13%
1.17	2.2%	VAT 17% (islands)
1.06	1.8%	VAT 6%

So roughly 82% of Διαύγεια amounts are VAT-inclusive, ~18% are not, and nothing in the record tells you which. Two consequences:

1. Summing amount_net across sources mixes units. For a consistent total use coalesce(amount_gross, amount_net) — a GROSS basis, which is the only one every source can support. That is what diavgis.gr now publishes, labelled «με ΦΠΑ».
2. Do not test Διαύγεια amounts against the Ν.4412 thresholds (€30,000 supplies/general services, €60,000 works and social/special services — both net of VAT). A €48,387 net procurement stored as its €60,000 gross reads as sitting exactly ON the ceiling. Threshold work must filter source = 'kimdis'.

TED values are net of VAT per Directive 2014/24/EU Art. 5(1); corroborated here on the small number of TED↔ΚΗΜΔΗΣ chains (8 of 11 match the ΚΗΜΔΗΣ net, none match the gross).

Caveats worth knowing

• Coverage is not 100% and the gap is not uniform. Against the sources over this window: TED ~98.9%, Διαύγεια procurement ~95.9%, ΚΗΜΔΗΣ ~80.9%. Treat counts as a large sample, not a census.
• Amounts have source typos. A handful of records carry absurd values. The live site caps aggregates at €5bn per record for this reason; do the same if you are ranking anything.
• If you use the OCDS feed for procurementMethod, note that limited covers direct awards. Counting a single bid as "single bidder" without excluding direct awards — where one bid is the definition — overstates the rate by roughly 16 percentage points. (bids_count is not in this file; the caveat is kept here because it is the mistake people make with this data, whichever route they take to it.)
• Chains are reconstructed, partly from explicit ΑΔΑ/ΑΔΑΜ references and partly by matching. method and confidence tell you which.

Format

Apache Parquet, zstd-compressed, one file per table. Read it with anything: pandas.read_parquet, polars.read_parquet, DuckDB (SELECT * FROM 'acts.parquet'), Spark, or the HuggingFace viewer above.

Two things worth knowing about the types:

• Identifiers are strings, not integers. A Greek ΑΦΜ can begin with 0 (373,321 buyer ΑΦΜs and 21,255 entity ΑΦΜs in this file do), so buyer_afm, supplier_afm, afm, cpv, nuts, id, ada and adam are all text. If you cast them to a number you will silently corrupt them.
• Nulls are real nulls, not empty strings — Parquet carries them natively.

Newlines inside text fields were collapsed to spaces during extraction, so titles and entity names are single-line even if you convert back to a line-oriented format.

Licence and attribution

CC-BY 4.0. Attribute as: Diavgis — Greek public procurement dataset, https://diavgis.gr

Underlying data is the open data of Διαύγεια, ΚΗΜΔΗΣ and EU TED. This release is the harvested, normalized and cross-linked derivative.

Live access

• Search and browse: https://diavgis.gr
• REST API and English documentation: https://diavgis.gr/docs/en
• Methodology and per-record provenance: https://diavgis.gr/about

Important

Diavgis is an independent private project. It is not the Greek government transparency portal diavgeia.gov.gr, is not ΚΗΜΔΗΣ or ΕΣΗΔΗΣ, and is not owned, operated or endorsed by the Greek State. It aggregates and cross-links the open data those official sources publish.