Skip to main content
This cookbook shows how to create a real ReAct-style agent eval with the TypeScript Evaris SDK. The example models a research agent used by support or operations teams during escalations. The agent needs to gather evidence from multiple sources before answering, so a ReAct runner with search and browser tools is a better fit than a single-shot generation flow. For the dataset, this recipe uses the Hugging Face ParthMandaliya/hotpot_qa distractor split. It is a strong stand-in for real research workflows because each sample usually requires multi-hop retrieval and answer synthesis rather than simple recall.

​
What this recipe does

  • Loads a Hugging Face dataset directly from the eval definition
  • Maps dataset fields from question and answer into Evaris input and target
  • Configures an inspect_native runner in react mode
  • Gives the agent web_search and web_browser tools
  • Uses the platform SDK auth flow so the client can create or reuse an agent eval suite automatically
  • Scores the run with lexical and judge-based scorers

​
Environment

Set these variables before running the script: 
export EVARIS_PLATFORM_URL=http://localhost:3000
export EVARIS_RUNTIME_URL=http://127.0.0.1:8100
export EVARIS_PROJECT_ID=proj_123
export EVARIS_API_KEY=evr_pk_xxx
export EVARIS_MODEL=openrouter/openai/gpt-4o-mini
export EVARIS_JUDGE_MODEL=openrouter/openai/gpt-4o-mini
export EVARIS_HOTPOT_LIMIT=25
export EVARIS_WEB_SEARCH_PROVIDER=tavily
export EVARIS_ENABLE_WEB_BROWSER=true
Optional: 
export EVARIS_SUITE_ID=es_123
If you omit EVARIS_WEB_SEARCH_PROVIDER, the example will skip the web_search tool. Set it to tavily, google, or exa only when that provider is configured in the runtime environment. If you omit EVARIS_ENABLE_WEB_BROWSER or set it to false, the example will skip the web_browser tool. Set it to true only when your sandbox image includes the Inspect web browser service. If EVARIS_SUITE_ID is omitted, the SDK will ensure a suite automatically when you use platformApiKeyAuth(...).

​
TypeScript example

The same example is available in the repo at examples/ts/react-hotpotqa-eval.mts. 
import { eval_, runtime } from "@evaris/sdk";

function requireEnv(name: string): string {
  const value = process.env[name]?.trim();
  if (!value) {
    throw new Error(`Missing required env var ${name}`);
  }
  return value;
}

function readPositiveInt(name: string, fallback: number): number {
  const raw = process.env[name]?.trim();
  if (!raw) {
    return fallback;
  }

  const value = Number.parseInt(raw, 10);
  if (!Number.isFinite(value) || value <= 0) {
    throw new Error(`${name} must be a positive integer`);
  }
  return value;
}

async function main(): Promise<void> {
  const runtimeBaseUrl =
    process.env.EVARIS_RUNTIME_URL?.trim() ?? "http://127.0.0.1:8100";
  const platformBaseUrl =
    process.env.EVARIS_PLATFORM_URL?.trim() ?? "http://localhost:3000";
  const projectId = requireEnv("EVARIS_PROJECT_ID");
  const apiKey = requireEnv("EVARIS_API_KEY");
  const suiteId = process.env.EVARIS_SUITE_ID?.trim();
  const agentModel =
    process.env.EVARIS_MODEL?.trim() ?? "openrouter/openai/gpt-4o-mini";
  const judgeModel =
    process.env.EVARIS_JUDGE_MODEL?.trim() ?? "openrouter/openai/gpt-4o-mini";
  const limit = readPositiveInt("EVARIS_HOTPOT_LIMIT", 25);
  const webSearchProvider = process.env.EVARIS_WEB_SEARCH_PROVIDER?.trim();
  const enableWebBrowser =
    (process.env.EVARIS_ENABLE_WEB_BROWSER?.trim().toLowerCase() ?? "") ===
    "true";
  const agentTools = [
    ...(webSearchProvider
      ? [eval_.tools.webSearch({ provider: webSearchProvider })]
      : []),
    ...(enableWebBrowser
      ? [eval_.tools.webBrowser({ interactive: false })]
      : []),
  ];

  const client = new runtime.Client({
    baseUrl: runtimeBaseUrl,
    auth: runtime.platformApiKeyAuth({
      platformBaseUrl,
      projectId,
      apiKey,
    }),
  });

  const evalDefinition = eval_.define({
    suite_id: suiteId,
    id: "react-support-research-hotpotqa",
    description:
      "ReAct research agent eval for support and operations escalations using HotpotQA.",
    data: eval_.datasets.huggingface("ParthMandaliya/hotpot_qa", {
      name: "distractor",
      split: "validation",
      limit,
      shuffle: true,
      seed: 17,
      sample_fields: {
        input: "question",
        target: "answer",
      },
    }),
    run: eval_.agentRunner({
      type: "inspect_native",
      mode: "react",
      model: {
        name: agentModel,
      },
      steps: [
        eval_.steps.systemMessage(
          [
            "You are a support escalation research agent.",
            "Investigate the question carefully before answering.",
            "Use tools when you need evidence.",
            "Return a concise final answer with no extra commentary.",
          ].join("\n"),
        ),
        eval_.steps.userMessage(
          [
            "Customer escalation question:",
            "{input}",
            "",
            "Work like a ReAct agent: search, inspect sources, then answer.",
          ].join("\n"),
        ),
        eval_.steps.agent(
          agentTools,
          {
            prompt:
              "Prefer a short search loop. Verify the answer before you submit it.",
            messageLimit: 12,
            maxAttempts: 2,
            truncation: "auto",
          },
        ),
      ],
      config: {
        use_case: "support-escalation-research",
      },
    }),
    score: [
      eval_.scorers.f1(),
      eval_.scorers.match({
        location: "any",
        ignoreCase: true,
        ignoreWhitespace: true,
        ignorePunctuation: true,
      }),
      eval_.scorers.modelGradedQa({
        model: judgeModel,
        partialCredit: true,
        instructions: [
          "Grade the answer on factual correctness.",
          "Reward answers that resolve the question directly.",
          "Penalize unsupported claims and missing key entities.",
        ].join(" "),
      }),
    ],
    channel: "sdk",
    labels: {
      cookbook: "react-agent",
      dataset: "hotpotqa",
      use_case: "support-research",
    },
    params: {
      cookbook: {
        scenario: "react-support-research",
        dataset: "ParthMandaliya/hotpot_qa",
        split: "validation",
      },
    },
  });

  const submitted = await client.submitEval(evalDefinition);
  const run = await client.waitForRun(submitted.job_id, {
    pollIntervalMs: 5_000,
    timeoutMs: 30 * 60 * 1_000,
  });

  console.log(
    JSON.stringify(
      {
        job_id: submitted.job_id,
        suite_id: run.suite_id,
        status: run.status,
      },
      null,
      2,
    ),
  );
}

main().catch((error: unknown) => {
  console.error(error);
  process.exitCode = 1;
});

​
Why this is a good ReAct eval

  • The task is retrieval-heavy and usually requires combining facts from multiple sources.
  • The agent has to decide when to search, when to open a source, and when to stop.
  • The scorer stack balances cheap lexical checks with a more realistic judge-based quality signal.
  • The suite type inferred by the SDK is agent, so this lands in the agent-eval path rather than a plain model eval path.

​
Running the example

Build the SDK first, then run the script with your usual TypeScript runner. If you want a fast smoke test, start with:
  • EVARIS_HOTPOT_LIMIT=5
  • a smaller agent model
  • the judge scorer removed or pointed at a cheaper model
After the run completes, inspect the run in Evaris to review:
  • per-sample answers
  • tool traces
  • score distributions
  • failure cases where the agent searched but still answered incorrectly