Built for addresses as people actually type them.
Typos, shortened street types, missing commas, the town written first: on input like this GridPin finds the right house more often than Nominatim and Photon.
Download free Run your first query How we measure
Fits: local, privacy-sensitive or high-volume address workloads in the United States (point addresses in 35 states, address ranges elsewhere), Canada, Germany (15 of 16 Länder; Bavaria not covered), France, Italy, Spain, Australia, Mexico, the Netherlands, Brazil and Serbia — forward, reverse and batch, address-first. Does not fit today: worldwide coverage, a hosted endpoint, or POI-first search outside France.
Found within 150 m of the true address
Share of all rows where the first answer lands within 150 m of an independent reference coordinate. Empty answers count as misses. Same queries, same truth, one run per engine.
| Query set | Rows | GridPin | Photon | Nominatim |
|---|---|---|---|---|
| Germany · distorted addresses | 2,000 | 82.1% | 46.9% | 15.8% |
| Germany · clean addresses | 2,000 | 82.8% | 83.2% | 81.2% |
| Spain · distorted addresses | 1,590 | 33.8% | 29.6% | 20.8% |
| Brazil · distorted addresses | 2,376 | 20.7% | 19.8% | 9.6% |
| Brazil · health-facility register (as typed) | 1,000 | 36.8% | 28.2% | 20.7% |
| United States · distorted addresses | 3,655 | 32.9% | 28.5% | 24.3% |
| United States · historic-places register (as recorded) | 1,000 | 64.1% | 38.9% | 52.3% |
| Canada · distorted addresses | 2,785 | 55.9% | 45.2% | 32.7% |
| Canada · radio-station sites (as recorded) | 1,000 | 39.3% | 39.3% | 25.2% |
| Mexico · distorted addresses | 2,000 | 24.1% | 13.8% | 7.7% |
| Mexico · cultural venues (as recorded) | 1,000 | 28.0% | 17.3% | 10.2% |
| Australia · distorted addresses | 2,000 | 74.6% | 50.6% | 42.2% |
| Australia · radio sites (as recorded) | 1,000 | 65.8% | 25.2% | 13.0% |
Distorted means each real address was changed by one fixed mechanical rule — a typo, a shortened street type, commas removed, the town moved to the front — with the reference coordinate kept byte for byte. On German and Spanish distorted addresses GridPin leads both services; on Brazilian distorted addresses GridPin and Photon are level and both are well ahead of Nominatim; on clean German addresses GridPin and Photon are level. We show the sets where GridPin leads or is level. On the other Spanish sets Photon leads, on clean Brazilian addresses Photon leads too (25.5% vs 23.2%); on clean US addresses Nominatim and Photon both lead (42.5% and 41.3% vs 39.1%); on clean Canadian addresses Photon leads (59.1% vs 57.5%; Nominatim 57.2%). On US distorted addresses and the historic-places register GridPin leads both services; on Canadian distorted addresses GridPin leads both, and on Canadian radio-station sites GridPin and Photon are level, both ahead of Nominatim; on Mexican distorted addresses, cultural venues and clean addresses GridPin leads both. On Australian distorted addresses, radio sites and clean addresses GridPin leads both. Full counts for every country, including sets where we trail, are below; this page makes no claim beyond the sets listed.
Coverage today includes the United States (point addresses in 35 states; address-range interpolation only in CO, FL, HI, MS, NH, NV and SC), Canada (all provinces and territories), Germany (15 of 16 Länder), France, Italy, Spain, Australia, Mexico, the Netherlands, Brazil and Serbia — not worldwide.
Details by country
For each country: what was measured, full counts per engine, significance and identity.
Germany · 2026-09-10 · 2 sets
What we tested. We took 2,000 real German addresses with known coordinates from two official registers — 1,000 public libraries and 1,000 electric-car charging stations — in 15 of 16 states (Bavaria is not covered). The rows were frozen before the run and never used while tuning the engine. Each address was tested twice: as written (clean) and with one mechanical distortion (dirty).
What it shows. Clean addresses: GridPin and Photon are level — the gap is within chance; GridPin is slightly ahead of Nominatim. Distorted addresses: GridPin finds 82% within 150 m, Photon 47%, Nominatim 16% — far beyond chance.
| Track | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| clean | gridpin | 2000 | 1958 | 42 | 1280 | 1656 | 1851 | 1891 |
| clean | nominatim | 2000 | 1913 | 87 | 1241 | 1623 | 1853 | 1904 |
| clean | photon | 2000 | 1975 | 25 | 1250 | 1663 | 1907 | 1965 |
| dirty | gridpin | 2000 | 1953 | 47 | 1267 | 1642 | 1836 | 1876 |
| dirty | nominatim | 2000 | 376 | 1624 | 256 | 317 | 366 | 374 |
| dirty | photon | 2000 | 1266 | 734 | 710 | 937 | 1074 | 1105 |
Technical detail for reproducing: significance test, data fingerprints
Germany, 2026-09-10 — 1,000 ISIL libraries and 1,000 BNetzA charging stations, 15 of 16 Länder (Bavaria not covered); rows sealed on 2026-09-09, outside the development regression set; source-family independence is not claimed.
Exact paired two-sided McNemar, ≤150 m (GridPin-only / comparator-only rows): clean vs Photon
p=0.6691367 (95 / 102) — not proof of parity; clean vs Nominatim p=0.0266263 (121 / 88);
dirty vs Photon p=1.65976e-155 (767 / 62), vs Nominatim p=7.50908e-366 (1,346 / 21).
Clean corpus SHA-256 9dd8e146b1c2fb90b028b0632b8a02d76c9273d821763ca01fff822fbc9a4f4b;
dirty 0b1719b75f59553dbb8ff897e784fe93d0fdb8b172bc5c0dd4576781f22c0674.
Spain · 2026-09-21 · 1 set
What we tested. 1,590 Spanish places from Wikidata with known coordinates; every address was changed by one mechanical distortion. One row lying outside Spain was dropped for all three engines.
What it shows. GridPin finds 33.8% within 150 m, Photon 29.6%, Nominatim 20.8%. GridPin’s lead over both is beyond chance, though smaller than in Germany. On Spain’s other test sets Photon leads.
| Set | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| distorted | gridpin | 1590 | 1048 | 542 | 385 | 538 | 738 | 958 |
| distorted | nominatim | 1590 | 511 | 1079 | 226 | 331 | 443 | 465 |
| distorted | photon | 1590 | 875 | 715 | 347 | 470 | 626 | 692 |
Technical detail for reproducing: significance test, data fingerprints
Spain, 2026-09-21 — 1,590 Spanish Wikidata items with a coordinate, each query
mechanically distorted; parent queries include locality descriptions, so this is not observed
human noise. One parent row outside the Spanish country contour (Overture 2026-08-19.0, land and maritime union) is excluded from all three engines; saved inputs and answers are unchanged. Release sheet spain.bin (v0.3.0)
c671b9a46089e32580189ee3c5a65e024779691a50ab26e77f5639023994819b, GridPin binary
9df585faa43b6766e29e14ea77b7153b3d7ae57d39665a5498b7eb8052e32f99; public Nominatim and a
local Photon (Spain extract).
Exact paired two-sided McNemar, ≤150 m: distorted vs Photon p=0.00344627 (297 / 229), vs Nominatim p=5.90642e-22 (340 / 133).
Brazil · 2026-10-06 · 3 sets
What we tested. Three sets with known coordinates: 2,376 places from Wikidata as written, the same 2,376 with one mechanical distortion each, and 1,000 health facilities from the national CNES register, spelled exactly as the register has them.
What it shows. Distorted addresses: GridPin and Photon are level (20.7% vs 19.8%, within chance), both well ahead of Nominatim (9.6%). Health-facility register: GridPin leads both (36.2% vs 28.2% and 20.7%), beyond chance. Clean addresses: Photon leads (25.5% vs 23.2%).
| Set | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| clean | gridpin | 2376 | 1551 | 825 | 270 | 551 | 972 | 1251 |
| clean | nominatim | 2376 | 1009 | 1367 | 172 | 412 | 783 | 884 |
| clean | photon | 2376 | 1494 | 882 | 311 | 607 | 1114 | 1271 |
| distorted | gridpin | 2376 | 1372 | 1004 | 229 | 492 | 846 | 1093 |
| distorted | nominatim | 2376 | 514 | 1862 | 100 | 229 | 413 | 461 |
| distorted | photon | 2376 | 1206 | 1170 | 241 | 471 | 879 | 996 |
| cnes | gridpin | 1000 | 752 | 248 | 270 | 368 | 604 | 707 |
| cnes | nominatim | 1000 | 512 | 488 | 122 | 207 | 449 | 504 |
| cnes | photon | 1000 | 802 | 198 | 160 | 282 | 643 | 756 |
Technical detail for reproducing: significance test, data fingerprints
Brazil, 2026-10-06 — three retained sets: clean Wikidata and mechanically distorted Wikidata (2,376 rows each) and CNES health-facility register as typed (1,000). Two clean rows outside the country contour are excluded equally for all three engines. Public Nominatim and public Photon answers were collected 2026-09-30 and retained; no new competitor requests were made. Top-1 great-circle distance to independent truth, computed with wikidata4_regress.py; empty answers remain in every denominator. These sets are not observed human search traffic.
Exact paired two-sided McNemar, ≤150 m: distorted vs Nominatim: p=1.72106e-34 (373 / 110); distorted vs Photon: p=0.386346 (277 / 256); cnes vs Nominatim: p=2.71682e-19 (246 / 85); cnes vs Photon: p=1.63145e-06 (202 / 116). These differences describe these sets only, not general superiority; the distorted Photon comparison is not proof of parity.
Measured engine SHA-256: 84bd75bc3264e71df57d3a76c9d69575ccaa24bcc4cd56975eb353830740cf4f; sheet br.bin SHA-256: 2d415c3a2abfd2df749b6a1d58625c12f8df1328a56248c9bc5eb5628cbc3ac7; release sheet brazil.bin (v0.4.0; metadata-only repack of the measured sheet, 2026-10-10) SHA-256: 016e102debb05aafad0ae409c3a2acec478b60bb48b74d03210b80dc52d42a6d. Corpora and raw receipts are retained evidence, not bundled in the public checkout.
United States · 2026-10-23 · 3 sets
What we tested. Three sets with known coordinates: 3,655 places from Wikidata as written, the same 3,655 with one mechanical distortion each, and 1,000 entries of the National Register of Historic Places spelled exactly as the register has them. Rows outside the 50 states and DC are excluded equally for all three engines. The reference coordinate is independent of the data GridPin ships.
What it shows. Distorted addresses: GridPin leads both (32.9% vs 28.5% and 24.3%), beyond chance. Historic-places register: GridPin leads both (64.1% vs 38.9% and 52.3%), beyond chance. Clean addresses: Nominatim and Photon both lead (42.5% and 41.3% vs 39.1%). One nationwide sheet with point addresses in 35 states and Census address ranges elsewhere; range answers are interpolated and land tens to hundreds of metres off.
| Set | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| clean | gridpin | 3655 | 3471 | 184 | 847 | 1430 | 2097 | 2292 |
| clean | nominatim | 3655 | 3222 | 433 | 991 | 1553 | 2192 | 2338 |
| clean | photon | 3655 | 3490 | 165 | 1010 | 1510 | 2224 | 2482 |
| distorted | gridpin | 3655 | 3063 | 592 | 721 | 1204 | 1700 | 1868 |
| distorted | nominatim | 3655 | 1796 | 1859 | 571 | 889 | 1330 | 1427 |
| distorted | photon | 3655 | 2864 | 791 | 698 | 1040 | 1601 | 1804 |
| nrhp | gridpin | 1000 | 895 | 105 | 468 | 641 | 763 | 825 |
| nrhp | nominatim | 1000 | 810 | 190 | 359 | 523 | 707 | 774 |
| nrhp | photon | 1000 | 880 | 120 | 271 | 389 | 642 | 783 |
Technical detail for reproducing: significance test, data fingerprints
United States, 2026-10-23 — three retained sets: clean Wikidata and mechanically distorted Wikidata (3,655 rows each) and the National Register of Historic Places as recorded (1,000). Public Nominatim and public Photon, answers collected 2026-10-20 and retained; no new competitor requests were made for the GridPin build measured on 2026-10-23.
Exact paired two-sided McNemar, ≤150 m: clean vs Nominatim: p=1.95005e-05 (348 / 471); clean vs Photon: p=0.013703 (474 / 554); distorted vs Nominatim: p=1.26392e-23 (657 / 342); distorted vs Photon: p=5.82017e-07 (616 / 452); nrhp vs Nominatim: p=5.22407e-15 (176 / 58); nrhp vs Photon: p=5.28401e-43 (309 / 57). These differences describe these sets only, not general superiority.
Measured engine SHA-256: f2dea2f0b9dc124b9342cb7c404b932c607af92bd77059ce9addb582d9bc5926; sheet us.bin SHA-256: 8fbdf76105057eab264dd40d37e080eeb506c09f23562718b9173035cae1a258; release sheet usa.bin (v0.5.0; metadata-only repack of the measured sheet, 2026-10-23): 4f3422f4d95939aefbcf454ceebc9a8dd16f78f2a13c5e6660028ddfad712cc0.
Canada · 2026-10-30 · 3 sets
What we tested. Three sets with known coordinates: 2,785 places from Wikidata as written, the same 2,785 with one mechanical distortion each, and 1,000 radio-station sites from the ISED Technical and Administrative Frequency List, spelled exactly as the licence records have them. Three Wikidata rows outside Canada are excluded equally for all three engines. The reference coordinate is independent of the data GridPin ships.
What it shows. Distorted addresses: GridPin leads both (55.9% vs 45.2% and 32.7%), beyond chance. Radio-station sites: GridPin and Photon are level (39.3% each; not proof of parity), both ahead of Nominatim (25.2%). Clean addresses: on clean Canadian addresses Photon leads (59.1% vs 57.5%; Nominatim 57.2%); this set cannot tell the three apart beyond chance. One nationwide sheet with point addresses in all provinces and territories; answers carry the Canadian postal code.
| Set | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| clean | gridpin | 2785 | 2698 | 87 | 1060 | 1602 | 1949 | 2045 |
| clean | nominatim | 2785 | 2587 | 198 | 1011 | 1593 | 1912 | 2035 |
| clean | photon | 2785 | 2751 | 34 | 1038 | 1646 | 1988 | 2110 |
| distorted | gridpin | 2785 | 2658 | 127 | 1032 | 1558 | 1897 | 1996 |
| distorted | nominatim | 2785 | 1493 | 1292 | 602 | 910 | 1109 | 1185 |
| distorted | photon | 2785 | 2453 | 332 | 827 | 1259 | 1535 | 1652 |
| tafl | gridpin | 1000 | 810 | 190 | 184 | 393 | 623 | 699 |
| tafl | nominatim | 1000 | 605 | 395 | 131 | 252 | 425 | 537 |
| tafl | photon | 1000 | 806 | 194 | 204 | 393 | 607 | 696 |
Technical detail for reproducing: significance test, data fingerprints
Canada, 2026-10-30 — three retained sets: clean Wikidata and mechanically distorted Wikidata (2,785 rows each) and ISED radio-station sites as recorded (1,000). Public Nominatim and public Photon, answers collected 2026-10-28 and retained; no new competitor requests were made for the GridPin build measured on 2026-10-30.
Exact paired two-sided McNemar, ≤150 m: clean vs Nominatim: p=0.774135 (393 / 384); clean vs Photon: p=0.114372 (349 / 393); distorted vs Nominatim: p=1.20016e-91 (867 / 219); distorted vs Photon: p=1.94838e-22 (625 / 326); tafl vs Nominatim: p=4.10126e-19 (200 / 59); tafl vs Photon: p=1 (120 / 120). These differences describe these sets only, not general superiority.
Measured engine SHA-256: f2dea2f0b9dc124b9342cb7c404b932c607af92bd77059ce9addb582d9bc5926; sheet ca.bin SHA-256: fcd4d9a9d1c5a1540722a66e7cad470ec515525f333df5f9ddfddea337449b44; release sheet canada.bin (v0.6.0; metadata-only repack of the measured sheet, 2026-10-30): 234515843c54b6c4feb93ab83c7b023a3e14beec8357c6d570eeed0f66f35400.
Mexico · 2026-11-04 · 3 sets
What we tested. Three sets with known coordinates: 2,000 clean addresses from two independent sources (909 places from Wikidata and 1,091 health facilities from the CLUES register of the Secretaría de Salud), the same 2,000 with one mechanical distortion each, and 1,000 cultural venues from the Sistema de Información Cultural (Secretaría de Cultura), addresses as the venues write them. Rows whose point lies outside Mexico were removed before selection. The reference coordinate is independent of the data GridPin ships.
What it shows. Distorted addresses: GridPin leads both (24.1% vs 13.8% and 7.7%), beyond chance. Cultural venues: GridPin leads both (28.0% vs 17.3% and 10.2%), beyond chance. Clean addresses: on clean Mexican addresses GridPin leads (29.2% vs Photon 15.9%, Nominatim 10.1%); the set has 2,000 rows, so the margin of error is about ±2 points. The address layer was surveyed in 2009–2010, so newer addresses are missing from it. One nationwide sheet; answers carry the postal code where the source has one.
| Set | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| clean | gridpin | 2000 | 1247 | 753 | 435 | 585 | 851 | 1012 |
| clean | nominatim | 2000 | 668 | 1332 | 110 | 202 | 416 | 462 |
| clean | photon | 2000 | 1046 | 954 | 186 | 318 | 697 | 818 |
| distorted | gridpin | 2000 | 1079 | 921 | 356 | 481 | 677 | 813 |
| distorted | nominatim | 2000 | 487 | 1513 | 84 | 153 | 302 | 340 |
| distorted | photon | 2000 | 1026 | 974 | 154 | 276 | 631 | 757 |
| sic | gridpin | 1000 | 616 | 384 | 200 | 280 | 422 | 533 |
| sic | nominatim | 1000 | 202 | 798 | 62 | 102 | 164 | 188 |
| sic | photon | 1000 | 406 | 594 | 102 | 173 | 324 | 370 |
Technical detail for reproducing: significance test, data fingerprints
Mexico, 2026-11-04 — three retained sets: clean addresses and mechanically distorted addresses (2,000 rows each; Wikidata and the CLUES health register) and cultural venues of the Sistema de Información Cultural as recorded (1,000). Public Nominatim and public Photon: answers for the clean and distorted sets collected 2026-11-04, for cultural venues 2026-10-31; all retained.
Exact paired two-sided McNemar, ≤150 m: clean vs Nominatim: p=2.99149e-64 (471 / 88); clean vs Photon: p=1.26171e-28 (431 / 164); distorted vs Nominatim: p=3.15518e-53 (411 / 83); distorted vs Photon: p=3.466e-18 (384 / 179); sic vs Nominatim: p=2.09717e-31 (216 / 38); sic vs Photon: p=3.80807e-10 (200 / 93). These differences describe these sets only, not general superiority.
Measured engine SHA-256: 09238e49275faf57e57790a77c10f5426bbef709c7b3893278194edb21483a03; release sheet mexico.bin (v0.7.0), on which the clean and distorted sets were measured: 6cc9e627bf84cf1126a5d31b71c14504cbbbf56eb5b1835731d85fc2a2ef454c; cultural venues on its pre-repack sheet mx.bin (metadata-only difference): c87ba719919e37ed80f8a9166d1b13f5e4be7e8a3ed272c923bc3bf653826110.
Australia · 2026-11-05 · 3 sets
What we tested. Three sets with known coordinates: 2,000 clean addresses (1,327 regular polling places of the 2025 federal election from the Australian Electoral Commission, 673 places from Wikidata), the same 2,000 with one mechanical distortion each, and 1,000 radio sites from the ACMA licence register, addresses as the licensees write them (New South Wales only). Wikidata rows whose coordinate cites a heritage register are excluded: those registers share their points with the address data.
What it shows. Distorted addresses: GridPin leads both (74.6% vs 50.6% and 42.2%), beyond chance. Radio sites: GridPin leads both (65.8% vs 25.2% and 13.0%), beyond chance. Clean addresses: GridPin leads both (75.0% vs 54.1% and 53.2%), on polling places 83.8% vs 59.0% and 59.6%, on Wikidata 57.5% vs 44.4% and 40.6%. Polling places are mostly schools and halls: the reference point is the building, the sheet's point is the address on the lot, so a correct answer can stand beyond 150 m. The AEC does not document how it places points; independence of the kept points was measured, not documented.
| Set | Engine | N | Found | Empty | ≤50 m | ≤150 m | ≤1,000 m | ≤5,000 m |
|---|---|---|---|---|---|---|---|---|
| clean | gridpin | 2000 | 1990 | 10 | 858 | 1499 | 1858 | 1925 |
| clean | nominatim | 2000 | 1888 | 112 | 494 | 1064 | 1673 | 1824 |
| clean | photon | 2000 | 1889 | 111 | 525 | 1082 | 1668 | 1828 |
| distorted | gridpin | 2000 | 1987 | 13 | 854 | 1492 | 1850 | 1921 |
| distorted | nominatim | 2000 | 1544 | 456 | 373 | 845 | 1358 | 1498 |
| distorted | photon | 2000 | 1817 | 183 | 479 | 1012 | 1574 | 1740 |
| acma | gridpin | 1000 | 932 | 68 | 430 | 658 | 857 | 911 |
| acma | nominatim | 1000 | 230 | 770 | 74 | 130 | 205 | 226 |
| acma | photon | 1000 | 404 | 596 | 168 | 252 | 354 | 376 |
Technical detail for reproducing: significance test, data fingerprints
Australia, 2026-11-05 — three retained sets: clean addresses and mechanically distorted addresses (2,000 rows each; AEC polling places and Wikidata) and ACMA radio sites as recorded (1,000). Public Nominatim and public Photon: answers collected 2026-11-04; all retained. Radio sites: Based on Australian Communications and Media Authority information.
Exact paired two-sided McNemar, ≤150 m: clean vs Nominatim: p=3.91875e-66 (562 / 127); clean vs Photon: p=1.73637e-60 (553 / 136); distorted vs Nominatim: p=1.08071e-126 (738 / 91); distorted vs Photon: p=5.71364e-74 (615 / 135); acma vs Nominatim: p=2.05511e-142 (540 / 12); acma vs Photon: p=1.55601e-90 (441 / 35). These differences describe these sets only, not general superiority.
Measured engine SHA-256: f2dea2f0b9dc124b9342cb7c404b932c607af92bd77059ce9addb582d9bc5926; release sheet australia.bin (v0.8.0): 6f19f33536894ff5b0d3444cee6d47ccc57a77a5170a3bcb01244114e68b3867; measured on its pre-repack sheet au.bin (metadata-only difference, answers byte-equal): 04690a2eacdafe663b949dc96f892d1694bf6942331a7199bd7119b9c1679bc5.
France, Italy, the Netherlands, Serbia — covered by the free sheets; no independent comparison with Photon and Nominatim is published for them on this page yet.
Full counts, method and data fingerprints; corpora and raw receipts are retained evidence, not bundled in the public source checkout.
Drop-in compatible
GridPin answers in the Photon API format, forward and reverse (the reverse endpoint is implemented
since 2026-08-06). On the external geocoder-tester suite — a compatibility fixture set, not an
accuracy oracle — the full official Alsace numbered-address file passes 230 / 285
(2026-08-07; 0 → 173 → 208 → 230 as the adapter matured), tested binary SHA-256
47a6ecf2cee6835dec7c7dd2fb331ebb5c8708f084e3b02f6bb5239b6e2103b8.
How we measure
- Independent truth. The reference coordinate never comes from the data GridPin ships; a set whose truth copies any engine's source is thrown out before comparing.
- Same input for everyone. Each engine gets the identical query and returns one answer. Nominatim is the public service, queried at most once per second; Photon is either the public service or a local instance of the country's OpenStreetMap extract — each country block says which.
- No small samples as results. Every set shown has at least 1,000 rows; empty answers stay in the denominator.
- Mechanical distortions, not guesses. Distortion rules are fixed before any engine answers and applied by a published hash, so no rule is tuned to a result.
Throughput, measured and stated with its conditions
How fast, honestly
Single arm64 laptop core, one thread, France sheet, warm mmap. Speed depends entirely on query shape: an exact match is an FST lookup, while a typo is answered by exploring edit-distance variants — hundreds of times more work.
| Query shape | Per core | Median latency |
|---|---|---|
| Exact, short (“10 rue de rivoli paris”) | ~3,000/s | <1 ms |
| Full clean address | ~600/s | <1 ms |
| Reverse (coordinate → address) | ~1,800/s | <1 ms |
| Typo / messy (fuzzy path) | ~10–50/s | ~60 ms |
Fuzzy matching is the slow path, and it is our headline feature — so we say so.
Feed messy input through geocode_many() or the batch command,
which spread work across every core. Making the fuzzy path itself faster is on the roadmap;
until it lands we would rather you size for tens of fuzzy queries per second per core than
be surprised in production.
Reproduce it yourself
The public contract runs without private sets or network services:
$ git clone https://github.com/gridpin/gridpin $ cd gridpin $ make public-bench-contract $ make smoke
The full record — every set above with its counts, paired tests, competitor versions and data fingerprints — is in docs-public/BENCHMARK.md.