vreau-digital/chatGPT/data-quality/geocoding-strategy-2026-05-11.md

Geocoding strategy — firms.entities

Data: 2026-05-11. Sub-agent A2.

Final coverage

Source	Rows	Accuracy	Notes
geonames_postal	2,128,990	~100m–2km	Exact 5/6-digit RO postal match against geonames RO.zip (firms.postal_codes).
photon	839,643	~50–500m	Komoot Photon OSM geocoder, free-text adr_full. Earlier batch (services/seap-scraper/src/geocode-photon.ts).
uat_centroid	670,657	5–30km	UAT polygon centroid match by locality+county.
judet_centroid	346,675	30–150km	Median of all postal codes within the judet. Filled the 2026-05-11 gap where judet_fallback was tagged but lat/lng never written.
seap_siruta_centroid	4,681	5–30km	NEW stub rows for SEAP-only CUIs (not present in ONRC firme dataset) using SIRUTA → gis_uats centroid.
seap_judet_centroid	2,497	30–150km	NEW stub rows for SEAP-only CUIs with city/county data in seap.cui_location.
unmapped	2	—	Two firms with literally zero address fields. Out of reach.

Total: 3,993,143 / 3,993,145 = 100.00 %.

Fallback chain (priority order)

For any new row entering firms.entities, apply in this order, stop at first hit:

1. Postal-code exact match → firms.postal_codes.postal_code = adr_cod_postal (5/6 digit). Source = geonames_postal.
2. Postal-code normalized (strip non-digit), same lookup. (Adds ~9K to the bucket — already covered in current dataset.)
3. Photon free-text on adr_full (OSM geocoder, requires network — see geocode-photon.ts).
4. UAT centroid by (adr_localitate, adr_judet) → firms.postal_codes median of matching place_name + county_code, OR public.gis_uats polygon centroid.
5. Judet centroid — median of all firms.postal_codes rows for the normalized judet name (upper(unaccent(replace(adr_judet,'MUNICIPIUL ','')))). 42 distinct judet keys cover all of RO + București.
6. SIRUTA centroid — for SEAP-mentioned CUIs only, where firms.entities row didn't exist: seap.announcements.{authority,supplier}_siruta → gis_uats.siruta centroid (transformed 3844→4326).
7. City+county from seap.cui_location → judet centroid fallback (seap_judet_centroid).

Authority / supplier coverage (downstream)

After backfill, JOIN-based coverage from SEAP:

Bucket	Total distinct CUIs	Geocoded	Pct
authority_cui	14,617	14,119	96.6 %
supplier_cui	65,675	64,793	98.7 %

Residual: 498 authorities + 882 suppliers (~1,373 unique) — these CUIs appear nowhere with address data (no siruta, no city/county in seap.cui_location, no usable address in any announcement). Most are malformed CUI strings (commas, semicolons, trailing punctuation) — should be cleaned up at SEAP ingestion. Out of scope for geocoding.

Cross-schema enrichment

• aaas.firme — 11 rows total, all 11 have geocoded parent in firms.entities via CUI. No action needed; UI agents JOIN.
• anre.licente — 27,275 rows with titular_cui populated, 11,043 distinct. All 11,043 CUIs match a geocoded firm. UI agents JOIN on firms.entities.cui = anre.licente.titular_cui.
• seap.announcements — supplier_address, authority_address, supplier_siruta, authority_siruta are populated. After this batch, almost every announcement can render on a map via firms.entities lookup.

Geom integrity

• firms.entities.geom (geography 4326) is now 1:1 with lat/lng (12,735 prior mismatches fixed where judet_fallback had stale geom from an older run).
• 2 unmapped firms have NULL on both. PostGIS spatial indexes still valid.

Forward maintenance

1. Anyone ingesting new firms (ANAF/ONRC weekly refresh) must apply the fallback chain in code before INSERT.
2. The seap_siruta_centroid and seap_judet_centroid stubs should be upgraded the moment an ANAF/ONRC record arrives for the same CUI — re-run the chain with the real adr_full.
3. If the SEAP CUI hygiene gets fixed (A1's domain), the 1,373 residual can be re-attempted.
4. judet_centroid (and the two seap variants) have only geocode_score = 0.1 and 0.3. UI clustering should down-weight or hide these at high zoom.

Queries used

All idempotent UPDATEs filtered on lat IS NULL. Centroid sources read from firms.postal_codes and public.gis_uats (SRID 3844 → 4326). Saved in-line in the agent transcript; the strategy itself is the artifact.