SafeRunner¶

SafeRunner wraps external inference or tool code so it passes through the same PolicyEnforcer entry points as the main RuntimeExecutor—input validation before the call, output validation after, with optional tracing.

Implementation: oris.integrations.safe_runner.

When to use it¶

You call Ollama, a REST API, or a vendor SDK outside the built-in component graph.
You still want consistent guard behavior and optionally a PipelineResult-shaped trace for observability.

Supported targets¶

A callable: (dict[str, Any]) -> Any
An object with run(self, dict[str, Any]) -> Any (preferred when both run and __call__ exist)

Return values are coerced to dict[str, Any] from:

dict / Mapping
Objects with a model_dump() that returns a mapping (e.g. Pydantic-style)

Example¶

from oris.integrations import SafeRunner
from oris.rai.policy import PolicyEnforcer

def my_external_llm(payload: dict) -> dict:
    q = payload.get("query", "")
    # ... call your model ...
    return {"output": f"stub response for: {q!r}"}

runner = SafeRunner(my_external_llm, policy=PolicyEnforcer())

plain = runner.run({"query": "Hello"})
traced = runner.run({"query": "Hello"}, include_trace=True)
print(traced.to_run_summary()["trace"][0]["flags"])

With include_trace=True, the trace includes a single external step with flags such as "kind": "external_pipeline".

Errors¶

GuardViolationError — Input or output failed policy (same as in-framework runs).
PipelineExecutionError — Invalid target, non-mapping output, or wrapped external failure (see source docstrings for semantics).

Notebook¶

The repository’s LLM integration notebook demonstrates optional Ollama + mock fallback behind SafeRunner.