mirror of
https://github.com/diegosouzapw/OmniRoute.git
synced 2026-07-31 20:32:20 +03:00
2 lines
1.9 KiB
JSON
2 lines
1.9 KiB
JSON
{"author":{"id":"MDQ6VXNlcjQwODM4MTI=","is_bot":false,"login":"apoapostolov","name":"Apostol Apostolov"},"body":"## Problem\nCombo building needs a quota-aware ranking rule that favors models whose remaining quota can last through the rest of the current quota window.\n\n## Proposed rule: Highest Remaining Quota Rate\nFor any service with a fixed quota window:\n\n- Let `remaining_quota` be the fraction of quota still available, normalized to `0..1`.\n- Let `remaining_days` be the time left in the current quota window, expressed in days.\n- Compute `remaining_quota_rate = remaining_quota / remaining_days`.\n\nFor weekly quotas:\n\n- The expected pacing threshold is `1/7` quota per day.\n- Prefer models where `remaining_quota_rate > 1/7`.\n- If `remaining_quota_rate <= 1/7`, stop using that model in active combos until it recovers above the threshold.\n\n## Example\nIf a model has `60%` of its weekly quota left and `2d 10hr` remaining:\n\n- `remaining_days = 2.4167`\n- `remaining_quota_rate = 0.60 / 2.4167 = 0.248/day`\n- Since `0.248 > 1/7`, the model should remain eligible and be preferred over lower-rate options.\n\n## Expected behavior\n- Combos should rank models using the `remaining_quota_rate` rule whenever a quota window is known.\n- Models above the threshold should be preferred or retained.\n- Models at or below the threshold should be dropped from active use.\n- The rule should be applied consistently during combo evaluation and refresh.\n\n## Notes\n- If a quota is measured in calls instead of percentage, normalize it to a fraction of the current window before applying the rule.\n- The same rule should generalize to other quota windows by using `1 / window_length_days` as the threshold.\n","comments":[],"createdAt":"2026-04-28T17:54:35Z","labels":[],"number":1735,"title":"[Feature] Combo Builder: quota-aware Highest Remaining Quota rule","url":"https://github.com/diegosouzapw/OmniRoute/issues/1735"}
|