Grading API now delivers sharper scores on an optimized engine

We optimized the inference engine behind the v5 graders: same models, sharper scores, at about the same speed and with the same API.

Action required

Now: none. GET /profile/grading and GET /job/grading keep their parameters, their algorithm keys and their response.
Later: none. Nothing in this update is deprecated.

What changed

grader-hrflow-profiles-cosmic5, grader-hrflow-profiles-titan5, grader-hrflow-jobs-cosmic5 and grader-hrflow-jobs-titan5 now run on an optimized inference engine: same models, sharper scores.
Most of their scores stay virtually the same; some are refined. If you filter profiles or jobs on a minimum score, check that threshold after the update.

Performance

Ranking: with grader-hrflow-profiles-titan5 on our evaluation set (38 jobs, 67 profiles each), each job's ranking agrees with the grading model's own ranking at 92.71% on average, up from 83.31% (Spearman rank correlation per job).
Speed: on our load test, speed is about the same as before.

Example

curl -G https://api.hrflow.ai/v1/profile/grading \
  -H "X-API-KEY: <your_api_key>" \
  -H "X-USER-EMAIL: <your_email>" \
  --data-urlencode "algorithm_key=grader-hrflow-profiles-cosmic5" \
  --data-urlencode "board_key=<board_key>" \
  --data-urlencode "job_key=<job_key>" \
  --data-urlencode "source_key=<source_key>" \
  --data-urlencode "profile_ids[0][source_key]=<source_key>" \
  --data-urlencode "profile_ids[0][profile_key]=<profile_key>"

The response keeps its shape: data.scores holds one [1-score, score] pair per profile, in the order of data.profiles.

{
  "code": 200,
  "message": "…",
  "data": {
    "profiles": [{"key": "<profile_key>"}],
    "scores": [[0.49184858798980713, 0.50815147161483765]],
    "axis": null,
    "errors": [null]
  }
}

Action required

What changed

Performance

Example

Links