Résumé parsing in 40% less time, tagging steady when busy
Résumé parsing now answers in 1.39 s instead of 2.33 s at the median, with the same parsed profile. POST /text/tagging stays fast when the service is busy: from 2 to 4 November no call took more than 5 s, against 1.35% of calls before, and 99 in 100 calls answered within 0.567 s instead of 7.06 s. Document OCR answers in 1.01 s instead of 1.35 s at the median, and text parsing no longer slows down when the service is busy.
Action required
- Now: none. Résumé parsing, Document OCR, text parsing and tagging keep their parameters and their responses.
- Later: none. Nothing in this update is deprecated.
What changed
- Résumé parsing: résumés sent to
POST /profile/parsing/fileare parsed faster, with every parsing model, and the parsed profile is the same as before. - Atlas and Chronos: résumés parsed with Atlas or Chronos save more than a second at the median (3.38 s instead of 4.45 s with Atlas, 5.72 s instead of 6.86 s with Chronos), most of all résumés with many experiences and educations.
- Tagging under load:
POST /text/taggingkeeps answering quickly when the service is busy, instead of waiting behind résumé parsings: no call took more than 5 s from 2 to 4 November, against 1.35% of calls before. A typical call was already fast and stays so: the median takes 0.072 s instead of 0.081 s. - Tagging by model: the slowest calls are faster with Programmatic Small and with the other Tagging Algorithms; Programmatic Large answers in the same time as before.
- Text parsing under load:
POST /text/parsingno longer waits behind résumé parsings when the service is busy: it keeps its usual response time. - Document OCR:
POST /text/ocranswers faster, with the same output.
Performance
| Time on our servers | Before | Now | Change |
|---|---|---|---|
| Résumé parsing, median | 2.33 s | 1.39 s | 40% less time |
| Résumé parsing, 99th percentile | 15.37 s | 8.04 s | 47% less time |
| Document OCR, median | 1.35 s | 1.01 s | 25% less time |
| Text parsing, 99th percentile | 10.74 s | 4.09 s | 2.6× faster |
| Tagging, median | 0.081 s | 0.072 s | 11% less time |
| Tagging, 99th percentile | 7.06 s | 0.567 s | 12.4× faster |
| Tagging, calls over 5 s | 1.35% | 0% | none left |
| Résumé parsing by model | Median, before | Median, now | 99th percentile, before | 99th percentile, now |
|---|---|---|---|---|
| Quicksilver | 1.82 s | 0.954 s | 12.82 s | 3.73 s |
| Hawk | 2.25 s | 1.54 s | 14.39 s | 3.94 s |
| Atlas | 4.45 s | 3.38 s | 16.92 s | 8.39 s |
| Chronos | 6.86 s | 5.72 s | no change | no change |
| Gaia | 8.75 s | 7.83 s | no change | no change |
| Tagging by model | Before | Now | Change |
|---|---|---|---|
| Programmatic Small, 99th percentile | 4.94 s | 0.944 s | 5.2× faster |
| Programmatic Large, median | 1.98 s | 2.2 s | no change |
| Other Tagging Algorithms, median | 0.081 s | 0.071 s | 12% less time |
| Other Tagging Algorithms, 99th percentile | 6.93 s | 0.253 s | 27.4× faster |
Measured on our servers for every successful request, at the same hours, on the five weekdays before 23 October 2026 and on 2, 3 and 4 November 2026. The time to send the call and its file to the API is not included. Programmatic Small and Programmatic Large are called on fewer days: their figures compare the October days with calls (8, 9 and 16 October for Small, 19 to 22 October for Large) with 2 to 4 November. "No change" means the new figure lies within the day-to-day spread of the old one. Tagging calls over 5 s are a share of every successful call of those days.
Example
curl -X POST https://api.hrflow.ai/v1/profile/parsing/file \
-H "X-API-KEY: <your_api_key>" \
-H "X-USER-EMAIL: <your_email>" \
-F "source_key=<source_key>" \
-F "file=@<resume.pdf>"The response keeps the format documented in the reference: the same parsed profile, sooner.