Conversation
…iably New tasks cover latency percentiles, trace-level annotation scores, time bucketing, annotation score filters, a single failing LLM span, sub-agent hierarchy, and a cost-to-input multi-hop question. The TRAIL loader posted annotations before Phoenix had ingested their spans, so each seed dropped a different subset. It now posts them after every span is queryable.
Contributor
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Automations to automatically generate PRs for you. |
This branch was successfully deployed
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Before: The TRAIL benchmark asked 10 questions about the research-assistant project. None of them covered latency, time windows, trace-level annotation scores, annotation score filters, span hierarchy, or questions that need several lookups chained together. Each fixture seed also silently dropped a different random subset of annotations (one seed kept 495 of 581 trail_errors, another kept 514).
After: The benchmark has 8 more tasks, ranging from easy to hard. A fresh seed keeps all 581 trail_error span annotations and all 585 trace scores.
Four of these (p95 LLM duration, 10-minute buckets, high-impact errors, average reliability) come from the questions in the Phoenix MCP SQL code-mode blog post that the benchmark did not cover.
How: Each task follows the existing layout and grader. Its
solve.pyreads Phoenix throughevals.harbor.verifiers.phoenix_api. The Python client cannot read trace annotations, so a newtrace_annotation_scoreshelper reads them through REST. Inscripts/load_patronus_trail.py, the loader now posts annotations after every span is queryable. Before, it posted them right after each trace's spans, and Phoenix dropped annotations whose span or trace it had not inserted yet.Validation:
-a oracle -e dockeron the 8 new tasks plus top-error-category, against a fixture reseeded with the fix: 9 of 9 trials scored 1.0.pytest tests/unit/harborpasses.Existing staged fixtures need
RESEED=1 make harbor-stagebefore avg-reliability and high-impact-traces will match.