Bring the stack you already run.

Takes in

DeepEval's saved test run (test_run_<time>.json, .latest_run_full.json, or the wrapped .latest_test_run.json when it holds results), one file per version; a metric name; DeepEval's own recorded threshold or a declared cut with its side; what identifies a case.

Gives out

A store run per version with one row per test case, met or missed, the score kept on the row; then fix-check and changes. Nothing imports DeepEval, no metric runs, no model is called, no prompt or answer text is kept.

Worked example

The HumanEval pair written through DeepEval 4.2.6 with a custom metric over stored test results (the files are DeepEval’s, the metric is ours, and the page says so first): 164 problems at two temperatures, 32 runs each, 10,496 test cases per file, one agent run twice 56 minutes apart with nothing changed. DeepEval reports 9,158 of 10,496 passed, then 9,109. fix-check: different, the new run failed more, p = 0.036, +0.47 points (+0.04 to +0.90); by input, p = 0.127; 2 of 328 cases flagged, both HumanEval/129.

Command

poisson-trace-import deepeval FILE [FILE ...] --store ROOT --run NAME
        --metric NAME
        --cut ( recorded | X ( --below | --above ) )
        --case ( name | input | position )
        [--started ISO-8601] [--json]

Not in it

Not one of DeepEval’s model-judged metrics; not Confident AI; not DeepEval’s SQLite store; not files from versions other than 4.2.6, which are untested.

The page

The full walkthrough, mirrored from the Trace repository.

Takes in

The score table you saved from Ragas (result.to_pandas() written as CSV, JSON Lines, JSON or Parquet, or an @experiment file), one or more files per version; one score column; a cut with its side; the column that names the case.

Gives out

A store run per version with one row per row of the table, met or missed; a row with no score has no outcome, is left out, and is counted in the report’s header. Then fix-check and changes. Nothing imports Ragas, no metric runs, no model is called, and nothing is kept of any column but the two you name.

Worked example

Two published result sets, a base model and the same model fine-tuned, on the same 150 rows of 146 questions, at a cut of 0.5 written down before any was run. Faithfulness (means 0.550, then 0.566): too small to tell, p = 0.629, could not have detected a drop in the share of misses below 15.9 points, and three repeats per question or 356 questions would settle it. Answer correctness: different, the new version failed less, p = 0.011, −13.3 points (−23.1 to −3.6). Answer relevancy: different, failed less, p = 0.009, −11.3 points (−19.3 to −3.4). Three scores at one cut are three tests. A second example shows a score Ragas could not give: 1,130 of 1,800 cases left out and counted, and no significant difference on the 670 left (p = 1.000, could not have detected a drop below 1.7 points).

Command

poisson-trace-import ragas FILE [FILE ...] --store ROOT --run NAME
        --metric COLUMN
        --cut X ( --below | --above )
        --case-column COLUMN
        [--started ISO-8601] [--json]

Not in it

It does not compare means; it compares the share of cases that missed the cut. With one run per case no single case can be flagged. Tables from Ragas versions other than 0.4.3 are read by their column names and no test runs another version; the version that scored the published pair is not stated.

The page

The full walkthrough, mirrored from the Trace repository.

Both instrumentors are needed. The CrewAI instrumentor alone stores the crew, the agent step and the tool and no model call; the OpenAI instrumentor alone stores the two model calls and nothing else, and the episode reads unanswered. Verified on reference agents; pinned versions are in the docs.

Infrastructure retries and a cost guard, landing the same rows in the same store as the local backend. The HumanEval pair on the Prove page ran on it: runs 20260905T185407Z-0313 and 20260905T195009Z-18ac.