Skip to content
Kindgi

Compare a version on a test set

A comparison runs an agent version on each case of a test set, as a replay of the past run, and rates its answers with the judgments people already made. You see, beside the recorded runs, whether people would rate the new version higher, before anyone uses it.

Compare a version

Section titled “Compare a version”

Version 0.1.0 of acme-ops.article-picker answered with three articles each time, and people said no to most of them. Version 0.2.0 changes its instructions:

instructions:
  "Call acme-ops.search-articles with the customer's question. Answer with only the articles that answer it, most relevant first: one is often enough.",

Under kindgi dev, saving the agent's file with version: '0.2.0' is enough: the test set keeps its copies of the 0.1.0 runs. On a runtime you deploy to, deploy the new version first.

Who may start one: an editor of the test set's project (or anyone who may execute the test set), who may write to the project the run lands in and run its agent or flow.

Terminal window
kindgi eval-runs start acme.article-picks --project=<project-id> --agent=acme-ops.article-picker --agent-version=0.2.0 --wait
{
  "runId": "b0b9f27a-14d9-4a4d-bc46-31aea6948d5f",
  "status": "completed",
  "result": {
    "summary": {
      "cases": 3,
      "candidate": { "kind": "agent", "agentId": "acme-ops.article-picker", "version": "0.2.0" },
      "baseline": { "kind": "recorded", "versions": [{ "agentId": "acme-ops.article-picker", "version": "0.1.0", "cases": 3 }] },
      "metrics": {
        "weightedYesShare": { "baseline": 0.3333333333333333, "candidate": 1, "delta": 0.6666666666666667, "n": 3, … },
        "judgedCoverage": { "baseline": 0.75, "candidate": 0.5, "delta": -0.25, "n": 3, … },
        "weightedPrecisionAtK": { "k": 10, "baseline": 0.3333333333333333, "candidate": 1, "delta": 0.6666666666666667, "n": 3, … }
      },
      "diverged": 1,
      "refusedWrites": 0,
      "errors": 0,
      "sampling": { "models": [{ "providerId": "anthropic", "model": "claude-sonnet-5-5", "runs": 3 }] },
      …
    },
    "perCase": [ … ]
  },
  …
}

Without --wait, the command returns the run's id at once, and kindgi eval-runs show <run-id> reads it later. From your code:

compare.ts
import { comparisonOf, createClient } from '@kindgi/sdk/client';


const kindgi = createClient(); // KINDGI_API_URL, KINDGI_API_TOKEN, or the running kindgi dev
const projectId = process.argv[2];


// Replay version 0.2.0 on each case of the test set.
const { runId } = await kindgi.evalRuns.start(
  'acme.article-picks',
  { agentRef: { agentId: 'acme-ops.article-picker', version: '0.2.0' } },
  { projectId },
);


let run = await kindgi.evalRuns.get(runId);
while (run.status === 'pending' || run.status === 'running') {
  await new Promise((resolve) => setTimeout(resolve, 2000));
  run = await kindgi.evalRuns.get(runId);
}


const comparison = comparisonOf(run); // undefined until it's finished
for (const [name, metric] of Object.entries(comparison?.summary.metrics ?? {})) {
  console.log(`${name}: ${metric.baseline} → ${metric.candidate}`);
}
judgedCoverage: 0.75 → 0.5
weightedYesShare: 0.3333333333333333 → 1
weightedPrecisionAtK: 0.3333333333333333 → 1

comparisonOf (Python: comparison_of) reads a finished comparison's result, typed: its summary and each case. It returns undefined (None) for a comparison that hasn't finished, or a dry run.

A model doesn't answer the same way every time: run the same comparison again, and the scores may differ. --repetitions=<n> (repetitions in the clients, up to 10) runs each case several times, and each metric gets a spread: its highest value minus its lowest.

Read the result

Section titled “Read the result”

Each metric has its value for the recorded runs (baseline) and for the version (candidate), the delta, and n, the cases it's based on.

  • weightedYesShare ("Judged yes" in the console): of the items in the answer that people judged, the share they said yes to, weighted by judge class. 0.1.0's answers had 3 yes of 9 judged articles; 0.2.0's had 3 of 3: it kept only the right article each time.
  • judgedCoverage ("Judged items"): the share of the answer's items that people judged. Items no one judged bring it down: here, each answer's text is an item too, and no one judged it.
  • weightedPrecisionAtK ("Top 10 judged yes"): judged yes among the first k items of a ranked list (10 unless you set --k).

How a judgment carries over. An item with its own id (an id or key field) is the same item when the id matches, at any place in the list: the yes on kb-301 counts for 0.2.0 too. The answer's text, and a list element without an id, carry a judgment only when they're unchanged; a reworded answer is a new item, with no judgments yet.

Count only restricted judges

Section titled “Count only restricted judges”

--class-weights=restricted-only (classWeights: 'restricted-only' in TypeScript, class_weights="restricted-only" in Python) counts only the judgments made under a restricted class; an item none of them judged counts as unjudged. The summary says which it used, in classWeights (as-recorded by default). On a test set of three cases, one of them judged only under a class restricted afterwards:

as-recorded restricted-only
weightedYesShare 0.556 (5 of 9) 0.5 (3 of 6)
judgedCoverage 1 0.667

Case by case, perCase has, for each case:

{
  "caseId": "240c0a0b-542b-4cb1-9156-fe7153a4377a",
  "runIds": ["732235ae-2321-46fe-876e-9e9818671cbf"],
  "changes": {
    "kept": [{ "key": "kb-101", "rankBefore": 0, "rank": 0 }],
    "dropped": [{ "key": "kb-102", "rankBefore": 1 }, { "key": "kb-103", "rankBefore": 2 }],
    "new": [{ "key": "answer", "pointer": "/appended/3/content" }]
  },
  "tools": [
    {
      "toolId": "acme-ops.search-articles",
      "toolVersion": "0.1.0",
      "arguments": { "query": "I forgot my password and can't sign in" },
      "source": "live",
      …
    }
  ],
  "diverged": true,
  "refusedWrites": 0,
  …
}
  • changes: the judged items the version kept (with their place before and now), the ones it dropped, and the new ones, for people to judge.
  • tools: each tool call, and what happened to it: live (it ran), recorded (it got the past run's result) or refused.
  • diverged: the version read something the past run didn't, so the read ran live. Here, 0.2.0 left the final period off the customer's question when it searched, so the call wasn't the same as 0.1.0's.

What a replay does

Section titled “What a replay does”

A replay does nothing the past run didn't do. "The same call" below means the same tool with the same arguments.

  • A read-only tool (mutating: false, no effect that changes things, and no approval to ask for) gets the past run's result for the same call, and runs live when the past run didn't make it (the case is then diverged). It also runs live when the version you compare pins other settings blocks than the past run's version, so a tool that reads its settings shows the change; that counts as diverged too. With --reads=live (reads: 'live' in the clients), it always runs live.
  • A read that reaches outside (a network or sensitive-data-egress effect) only ever gets the past run's result for the same call.
  • Any other tool gets the past run's result for the same call, or doesn't run: the model is told replay: this call changes things and has no recorded result (or that the call needs an approval, or reaches outside), and the case shows what it would have done. refusedWrites counts these calls.
  • A replay never waits for an approval. Where the past run's conversation was approved or rejected, the replay follows that decision; where there's no decision, the case says approvalSkipped.
  • A replay runs for the past turn's end user. An agent whose memory is per end user (same-user) is offered the same tools as the past turn, kindgi_remember included, which stores nothing in a replay. A replay's conversation is never recalled later as an earlier conversation.

The replay runs are ordinary runs, marked with replayOf (the past run) and evalRunId. Lists of runs leave them out: kindgi runs list --eval-run=<run-id> lists one comparison's, and --replays=include or --replays=only adds them to any list. Conversation lists leave out their conversations the same way (kindgi conversations list --replays=only). You can judge them like any run: that's how a version's new items get their judgments.

A comparison stopped before it finished, because the runtime stopped, ends failed with the cases it got through: Interrupted: the eval run's executor stopped before it finished. 2 of 3 cases finished. Start it again.

Compare a flow version

Section titled “Compare a flow version”

A test set built from a flow's judged runs (kindgi eval-suites from-judgments … --flow=<flow-id>) compares a flow version. Each case runs the past run's input through that version, as a replay, and the flow's whole output is rated: the output itself (the output item) and each element of a list in it.

To try some of its tools or agents at another version, without publishing a new flow version, add --with=<id>@<version> (once for each). Here the flow acme-ops.suggest-articles 0.1.0 runs its acme-ops.find-articles tool at 0.2.0, which matches only the articles a question asks about:

Terminal window
kindgi eval-runs start acme.flow-picks --project=<project-id> --flow=acme-ops.suggest-articles --flow-version=0.1.0 --with=acme-ops.find-articles@0.2.0 --reads=live --wait
"summary": {
  "candidate": {
    "kind": "flow",
    "flowId": "acme-ops.suggest-articles",
    "version": "0.1.0",
    "versions": { "tools": { "acme-ops.find-articles": "0.2.0" } }
  },
  "metrics": {
    "weightedYesShare": { "baseline": 0.5, "candidate": 1, … },
    …
  },
  …
}
  • In the clients: flowRef: { flowId, version } in place of agentRef, and versions: { tools: { 'acme-ops.find-articles': '0.2.0' } } (agents under agents).
  • A read-only tool's new version shows only with live reads. With the default, recorded reads, its call gets the past run's result: the same comparison without --reads=live rates 50% before and after.
  • Under kindgi dev, a pack serves two versions of a tool side by side, so a tool swap works there. An agent has one version there, its file's, so swapping an agent needs a runtime you deploy to.
  • A --with the flow doesn't use is refused: flow acme-ops.suggest-articles 0.1.0 doesn't use acme-ops.article-picker.

A flow stops at a write the replay refuses. acme-ops.suggest-and-log also records the articles it suggested. With 0.2.0 of the tool, it would have recorded other articles than the past run did for one question, so that case stops there:

"stopped": {
  "toolId": "acme-ops.log-suggestion",
  "arguments": {
    "articles": [{ "id": "kb-201", "title": "Request a refund" }],
    "question": "I was charged twice for one order. How do I get my money back?"
  },
  "reason": "replay: this call changes things and has no recorded result"
}

A stopped case isn't an error: nothing was written, and it has no output to rate, so it's left out of the metrics. The summary counts it in stopped (and the refused call in refusedWrites); the other cases are rated as usual. A step's agent turn doesn't stop the flow: its refused calls go back to the model, as in an agent comparison.

When a version does better, make it live where you choose: see Release an agent version.

In the console

Section titled “In the console”

A test set's page has Compare an agent version: pick the version, how many runs per case, whether to read live data instead of the past runs' results, and Count only restricted judge classes. Its Comparisons list links each one to its page, which has:

  • a headline in words: "People's judgments rate acme-ops.article-picker 0.2.0 better than the recorded runs: 33% yes before, 100% now, on the 3 of 3 cases with answers people judged. In 1 case it read something the recording didn't have.";
  • the metrics, for the recorded runs and the version, with the change (a comparison that counted only restricted classes says so: "Only judgments in judge classes restricted to some judges counted; the others weighed 0.", and lists mark it "restricted classes only");
  • Case by case: the recorded answer beside the version's, the judged items it kept and the ones no longer there, the new ones, and each tool call;
  • The raw result, as JSON.

A flow's test set has Compare a flow version instead, with a picker for each agent its steps run, to run it at another version; tools are swapped from the CLI and the clients. A flow comparison's page shows the recorded output beside the version's, and a stopped case says what it would have done: "It stopped before a write the recorded run didn't make: it would have called acme-ops.log-suggestion with {"articles":[…],"question":"…"}. Nothing was written, and there's no output to rate."