Free-form Indian addresses can combine sub-premise notation, transliterated locality names, landmarks, inconsistent component order, and missing or incorrect PIN codes. A high match rate alone can hide systematic positional error—especially if locality or PIN centroids are counted as successful matches.
I’m using the following working checklist in a civic/public-data workflow:
Keep raw and normalized values. Never overwrite source text. Parse a copy with tools such as libpostal plus local rules; record language/script and landmark text.
Generate and retain multiple candidates. Compare structured and free-form Nominatim queries. Save candidate rank, match granularity, provider/reference-data version, and timestamp.
Treat PIN and locality as validation evidence, not automatic truth. Disagreements should trigger review. Document whether each field constrained the search or merely flagged risk.
Report match rate and positional accuracy separately. Stratify labeled samples by geography and address type. Distinguish rooftop, street, locality, and PIN-centroid matches, and manually review low-confidence cases.
Stress-test downstream spatial analysis. Flag or exclude centroid-only points where appropriate, use suitable distance metrics or projected coordinate systems for clustering, and test whether geocoding uncertainty changes the clusters. The HDBSCAN methodology is useful for understanding density, noise, and cluster stability.
Treat DIGIPIN as complementary. India Post’s DIGIPIN encodes a location, but does not by itself validate a postal address or match a household.
What validation signals have proved most useful in your own Indian-address pipelines?