fraud_hunter_env / client.py
VanshGoel1's picture
Upload folder using huggingface_hub
a3ad3e6 verified
Raw
History Blame Contribute Delete
2.19 kB
"""Fraud Hunter Env — EnvClient subclass for training-side use."""
from __future__ import annotations
from typing import Any, Dict
from openenv.core import EnvClient
from openenv.core.client_types import StepResult
from openenv.core.env_server.types import State
from .models import FraudHunterAction, FraudHunterObservation
class FraudHunterEnv(EnvClient[FraudHunterAction, FraudHunterObservation, State]):
"""
WebSocket client for the Fraud Hunter Env server.
Example:
>>> with FraudHunterEnv(base_url="http://localhost:8000") as env:
... r = env.reset()
... r = env.step(FraudHunterAction(kind="query_corporate", entity_name="Acme"))
... print(r.reward, r.observation.tool_output)
"""
def _step_payload(self, action: FraudHunterAction) -> Dict[str, Any]:
# Pydantic v2 serialisation, dropping None fields for compact WS frames.
return action.model_dump(exclude_none=True, mode="json")
def _parse_result(self, payload: Dict[str, Any]) -> StepResult[FraudHunterObservation]:
obs_data = payload.get("observation", {}) or {}
observation = FraudHunterObservation(
case_brief=obs_data.get("case_brief"),
tool_output=obs_data.get("tool_output"),
base64_document=obs_data.get("base64_document"),
grader_feedback=obs_data.get("grader_feedback"),
evidence_graph=obs_data.get("evidence_graph"),
step_count=obs_data.get("step_count", 0),
budget_remaining=obs_data.get("budget_remaining", 0),
difficulty_tier=obs_data.get("difficulty_tier", 1),
info=obs_data.get("info"),
done=payload.get("done", False),
reward=payload.get("reward"),
)
return StepResult(
observation=observation,
reward=payload.get("reward"),
done=payload.get("done", False),
)
def _parse_state(self, payload: Dict[str, Any]) -> State:
return State(
episode_id=payload.get("episode_id"),
step_count=payload.get("step_count", 0),
)