Skip to content

fix(trace): 评估运行的 trace input 不再回退成 answer #161

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wxrbyte wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(trace): 评估运行的 trace input 不再回退成 answer #161

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
45 changes: 45 additions & 0 deletions packages/core/src/trace-projection.test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,4 +224,49 @@ describe('projectTrace (V9.1)', () => {
const toolStep = trace.steps.find((s) => s.kind === 'tool' && s.tool?.id === 'live-run');
expect(toolStep?.status).toBe('running');
});

it('projects the benchmark case prompt as the trace input, never the answer', () => {
const caseDef: EvaluationCase = {
id: 'case-1',
name: 'NVDA research',
category: 'tool-selection',
difficulty: 'golden',
input: { prompt: 'Research NVDA' },
expected: {},
tags: [],
source: 'hand-authored',
};
const evalRun: EvaluationRun = {
id: 'eval-run-1',
experimentId: 'exp-1',
caseId: 'case-1',
datasetId: 'ds-1',
status: 'completed',
startedAt: 1000,
completedAt: 9000,
answer: 'NVDA looks fine',
toolCalls: [],
failureModes: [],
};
const trace = projectTrace({ evaluationRun: evalRun, evaluationCase: caseDef });
expect(trace.input).toBe('Research NVDA');
expect(trace.input).not.toBe(evalRun.answer);
});

it('leaves the trace input empty when no authoritative prompt was recorded', () => {
const evalRun: EvaluationRun = {
id: 'eval-run-2',
experimentId: 'exp-1',
caseId: 'case-2',
datasetId: 'ds-1',
status: 'completed',
startedAt: 1000,
completedAt: 9000,
answer: 'NVDA looks fine',
toolCalls: [],
failureModes: [],
};
const trace = projectTrace({ evaluationRun: evalRun });
expect(trace.input).toBe('');
});
});
6 changes: 5 additions & 1 deletion packages/core/src/trace-projection.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -291,7 +291,11 @@ export function projectTrace(input: TraceProjectionInput): FolioTrace {
startedAt,
completedAt,
latencyMs,
input: run?.input ?? evaluationRun?.answer ?? '',
// The evaluation run record carries no input of its own; the benchmark
// case prompt is the authoritative input for that run (spec §9). Falling
// back to `answer` would project the OUTPUT as the input, which is the
// exact reconstruction-from-run-state the projection must never do.
input: run?.input ?? evaluationCase?.input.prompt ?? '',
answer,
error: run?.error?.message ?? evaluationRun?.error?.message,
completeness: deriveCompleteness(Boolean(run || evaluationRun), tools, traceEvents, traceRefForOutput),
Expand Down
Loading