- Report
- FR-009
- Published
- Written by
- Anders Holm
- Reading time
- 02 min / 379 words
Thirty years of drawings, one query
How we indexed 1.9 million drawings from a hospital estate so that any one of them can be found in eleven milliseconds.
Section 01
An inherited archive
When a regional hospital estate asked us to take over the maintenance of its building records, the archive arrived on 214 hard drives, eleven filing cabinets and a microfilm reader. It covered thirty years and forty-one buildings: about 1.9 million drawings, specifications and survey files in more than sixty formats, from scanned blueprints to the native files of software that no longer runs. Finding the current drawing for a single plant room took the estate team an average of two days. Most of the time they gave up and measured the room again.
Section 02
Indexing what was never indexed
We converted every file to an open format and read the title blocks with text recognition, checked by a person wherever the confidence was low. Each drawing was then registered to the coordinate grid of the estate, so it can be found by location as well as by name: select a room in the model and every drawing that has ever shown it appears in order. The index sits on top of the original files rather than replacing them. As the archive grew past a million records, query time at the 95th percentile fell from seconds to eleven milliseconds.
Fig. 009.1/ Query time, 95th percentile
Unit / ms
| Period | P95 |
|---|---|
| Jan 24 | 2 400 ms |
| Mar 24 | 1 350 ms |
| May 24 | 620 ms |
| Jul 24 | 240 ms |
| Sep 24 | 95 ms |
| Nov 24 | 41 ms |
| Jan 25 | 22 ms |
| Mar 25 | 14 ms |
| May 25 | 11 ms |
Section 03
Readable in fifty years
Speed is the visible part, but the harder requirement is time. Every record is stored as PDF/A, IFC or E57, all open and documented formats, and written to three sites in two countries with a checksum verified every week. If our own software disappeared tomorrow, the archive would still be readable with tools that exist today. The engineers on the estate now find any drawing in about the time it takes to ask for it. More importantly, so will whoever maintains these buildings in 2075.
Section 04
What an index costs
Indexing 1.9 million records took fourteen months and a team of four, most of it spent on the oldest material: blueprints from the 1990s scanned at low resolution, where text recognition needed a person to check one title block in five. The newest files were hard in a different way, because they arrive in proprietary formats that change every few years. Converting at the door, on the day a file enters the archive, now costs a few seconds per drawing. Converting thirty years later cost the estate a year. The cheapest index is the one built while the files are still young.
Fig. 009.2/ Records indexed by decade of origin
- 1990s212 000
- 2000s486 000
- 2010s731 000
- 2020s471 000
