Repository navigation
Fix generation quota pacing and withholding failures - #2
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A ten-question tuning audit exposed empty withholding responses being treated as errors and frequent rate-limit failures. Q008 diagnostics also showed a verifier returning supported=false while confirming that the feature was not supported.
Generation now paces calls using token/reset headers, buffers bounded retries and preserves request deadlines. Empty non-answers are sanitized; released answers still require nonempty text and valid citations. The verifier explicitly distinguishes evidence support from product feature support, and the prompt/cache version advances. Offline collection retains private stage results and content-free HTTP attempt statuses; API responses do not expose drafts.
Validation: 109 clean-public-checkout tests pass; local application checks 123 passed/1 skipped, and 16 focused evaluation checks pass (overlapping sets). Three saved-evidence diagnostic reruns completed with 8/8 HTTP 200 responses. After the verifier wording change, a separate Q008 check released a cited answer with positive verification, using 3/3 HTTP 200 responses. These are targeted checks, not a ten-case quality improvement benchmark. No document re-embedding or new live RAGAS calls.
Quota estimates are approximate and local to one provider instance; sustained multi-worker traffic needs shared quota coordination. Private traces and research notes remain local.