Parameter Injection

Experiment parameters flow through request_mapping for both REST and SDK connector endpoints. Use \{\{ experiment_parameters.model \}\} (or the deprecated alias \{\{ params.model \}\}) in your templates.

How it works

When a test run is associated with an experiment, the platform resolves the experiment’s parameter values once at queue time and stores an immutable snapshot on the test run. During execution, the snapshot is injected into the Jinja2 template context as experiment_parameters (and the deprecated alias params).

Your request mapping references individual values with dot notation:

{
"model": "{{ experiment_parameters.model }}",
"temperature": {{ experiment_parameters.temperature }},
"messages": [{"role": "user", "content": "{{ input }}"}]
}

This works identically whether the endpoint is REST (HTTP) or SDK connector (WebSocket). The template is rendered before the request is dispatched, so your API or function receives the final resolved values.

REST endpoints

For REST endpoints, the rendered request body is sent as the HTTP payload. See Using Experiment Parameters for full details, including how to use Jinja2 defaults for graceful fallback when no experiment is attached.

SDK connector endpoints

For SDK connector endpoints, the rendered request mapping becomes the function’s keyword arguments. The function signature simply declares the parameters it expects:

from rhesis.sdk import endpoint

@endpoint(
    name="rag",
    request_mapping={
        "query": "{{ input }}",
        "model": "{{ experiment_parameters.model | default('gpt-4') }}",
        "temperature": "{{ experiment_parameters.temperature | default(0.7) }}",
        "system_prompt": "{{ experiment_parameters.system_prompt | default(none) }}",
    },
    response_mapping={
        "output": "{{ response }}",
    },
)
def rag(query: str, *, model: str, temperature: float, system_prompt: str = None):
    return {"response": llm.chat(
        model=model,
        temperature=temperature,
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": query},
        ],
    )}

During a test run with an experiment, experiment_parameters.model resolves to the experiment’s value (e.g., "gpt-4o"). Without an experiment, the Jinja2 default() filter provides the fallback.

There is no distinction between “test data” and “configuration” at the request mapping level — both are template variables rendered the same way. This keeps the mental model simple: input is the test prompt, experiment_parameters.* are the experiment knobs, and any other platform variables (test_id, session_id, etc.) work as documented in Platform-Managed Variables.

Parameters.get() in production code

The experiment_parameters template variable handles test-time parameter injection. For production traffic (live requests that don’t go through the test runner), your application reads parameters using the Parameters.get() facade:

from rhesis.sdk import Parameters

params = Parameters.get("Customer Support", environment="default")
model = params.model
temperature = params.get("temperature", 0.7)

During a test run, the platform also sets an internal context variable so that Parameters.get() calls inside your function return the experiment’s values without hitting the network. This means production code that calls Parameters.get() will automatically pick up the experiment snapshot during tests — no special branching needed.

See SDK Usage for the full Parameters.get() API.

The Wire Protocol

For those debugging at the protocol level, the ExecuteTestMessage payload sent by the Rhesis platform over the connector includes the full parameter snapshot:

• experiment_parameters / parameters: The resolved Dict[str, Any] of experiment parameter values. Both field names are sent; experiment_parameters is preferred.
• test_parameters: The test’s freeform Dict[str, Any] of per-test data (see Test Parameters).
• parameter_version: The sequential version identifier (e.g. v3) of the version being executed.
• parameter_experiment_id: The UUID of the experiment.
• parameter_source: The resolution method used ("environment", "experiment_id", or "version"). Older snapshots may still show "label"; treat it as "environment".
• parameter_source_environment: The environment name when resolution went through an environment (for example "default"). Legacy snapshots may use parameter_source_label instead.
• parameter_schema: The full schema, allowing the SDK to validate the payload locally.

Because the backend dispatcher copies parameter_source and parameter_source_environment from the run snapshot, the ResolvedParameters.source_environment inside your application reflects the environment (for example “via default”) that was targeted at run-queue time.

Worked Example: The Chatbot

The Rhesis chatbot demo application is the canonical reference for parameter management.

The core chat() function is decorated as an @endpoint. Parameters come through the request mapping alongside test data:

@endpoint(
    name="chat",
    request_mapping={
        "message": "{{ input }}",
        "model": "{{ experiment_parameters.model | default(none) }}",
        "temperature": "{{ experiment_parameters.temperature | default(0.7) }}",
        "system_prompt": "{{ experiment_parameters.system_prompt | default(none) }}",
        "use_case": "{{ experiment_parameters.use_case | default('travel') }}",
    },
    response_mapping={
        "output": "{{ message }}",
        "session_id": "{{ session_id }}",
        "context": "{{ context }}",
    },
)
async def chat(
    message: str,
    *,
    model: Optional[str] = None,
    temperature: float = 0.7,
    system_prompt: Optional[str] = None,
    use_case: str = "travel",
) -> ChatResponse:
    # model, temperature, system_prompt come from the experiment
    # during test runs, or from defaults otherwise
    ...

The FastAPI route serves production traffic and resolves parameters through the Parameters.get() facade, then calls the same function:

def _resolve_chatbot_params() -> dict:
    try:
        params = Parameters.get("Chatbot Demo", environment="default")
        return {
            "model": params.model,
            "temperature": params.get("temperature", 0.7),
        }
    except Exception:
        return {"model": os.getenv("DEFAULT_GENERATION_MODEL"), "temperature": 0.7}

@app.post("/chat")
async def chat_endpoint(request: Request, chat_request: ChatRequest):
    params = _resolve_chatbot_params()
    result = await chat(
        message=chat_request.message,
        model=params.get("model"),
        temperature=params.get("temperature"),
    )
    return result

Test Parameters vs Experiment Parameters

The template context exposes two separate namespaces for injected data:

Use test_parameters for per-test data that varies across tests in the same run (e.g. a customer ID, a locale, or a scenario-specific flag). Use params for configuration that applies uniformly to every test in a run (e.g. model name, temperature, system prompt).

Both namespaces are available in the same request mapping and do not collide:

{
  "model": "{{ params.model }}",
  "messages": [
    {
      "role": "user",
      "content": "{{ input }}"
    }
  ],
  "customer_id": "{{ test_parameters.customer_id }}"
}