File size: 10,798 Bytes
61f1214 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | """
AI Interview Simulator — OpenEnv Environment
============================================
Implements the full OpenEnv interface:
reset() → InterviewObservation
step() → StepResult(observation, reward, done, info)
state() → InterviewState
The environment simulates a real job interview: the agent receives one
question at a time and must provide a text response. After every answer
a deterministic grader scores it (0.0–1.0) and issues partial reward.
The episode ends when all questions have been answered.
Usage:
env = InterviewEnv(task_id="junior_behavioral")
obs = env.reset()
while True:
action = InterviewAction(response="My answer …")
result = env.step(action)
if result.done:
break
"""
from __future__ import annotations
import time
from typing import Any, Dict, Optional
from interview_env.graders import grade_response
from interview_env.models import (
DifficultyLevel,
InterviewAction,
InterviewObservation,
InterviewStage,
InterviewState,
QAPair,
StepResult,
)
from interview_env.tasks import TASK_REGISTRY, TaskDefinition, get_task
class InterviewEnv:
"""
OpenEnv-compliant AI Interview Simulator environment.
Parameters
----------
task_id : str
One of: "junior_behavioral", "mid_technical", "senior_system_design"
max_response_length : int
Characters — responses exceeding this are truncated (soft limit).
penalty_per_truncation : float
Reward deducted when the agent's response is truncated.
"""
VALID_TASKS = list(TASK_REGISTRY.keys())
DEFAULT_TASK = "junior_behavioral"
def __init__(
self,
task_id: str = DEFAULT_TASK,
max_response_length: int = 2000,
penalty_per_truncation: float = 0.05,
) -> None:
if task_id not in TASK_REGISTRY:
raise ValueError(
f"Unknown task_id '{task_id}'. "
f"Choose from: {self.VALID_TASKS}"
)
self._task_id = task_id
self._max_response_length = max_response_length
self._penalty_per_truncation = penalty_per_truncation
# Internal state (initialised by reset)
self._task: TaskDefinition = get_task(task_id)
self._current_index: int = 0
self._qa_history: list[QAPair] = []
self._per_q_scores: list[float] = []
self._done: bool = False
self._total_steps: int = 0
self._start_time: float = time.time()
self._cumulative_reward: float = 0.0
# ──────────────────────────────────────────
# OpenEnv Interface
# ──────────────────────────────────────────
def reset(self) -> InterviewObservation:
"""
Reset the environment to the beginning of the interview.
Returns the first question as an observation.
"""
self._task = get_task(self._task_id)
self._current_index = 0
self._qa_history = []
self._per_q_scores = []
self._done = False
self._total_steps = 0
self._start_time = time.time()
self._cumulative_reward = 0.0
return self._build_observation()
def step(self, action: InterviewAction) -> StepResult:
"""
Submit the agent's answer to the current interview question.
Returns
-------
StepResult with:
observation — next question (or completion message if done)
reward — per-step score in [–0.05, 1.0]
done — True when all questions are answered
info — detailed grader breakdown
"""
if self._done:
# No-op after episode ends
return StepResult(
observation=self._build_observation(),
reward=0.0,
done=True,
info={"warning": "Episode already complete. Call reset()."},
)
self._total_steps += 1
current_q = self._task.questions[self._current_index]
# Soft-truncate response and penalise
response = action.response
truncated = False
reward_adj = 0.0
if len(response) > self._max_response_length:
response = response[: self._max_response_length]
truncated = True
reward_adj = -self._penalty_per_truncation
# Grade the response
reward_obj = grade_response(response, current_q)
step_reward = reward_obj.total_score + reward_adj
step_reward = max(0.0, min(1.0, step_reward))
# Record in history
qa = QAPair(
question = current_q.text,
answer = response,
stage = current_q.stage,
question_score = reward_obj.total_score,
)
self._qa_history.append(qa)
self._per_q_scores.append(reward_obj.total_score)
self._cumulative_reward += step_reward
# Advance to next question
self._current_index += 1
if self._current_index >= len(self._task.questions):
self._done = True
obs = self._build_observation()
info: Dict[str, Any] = {
"question_answered": current_q.text,
"grader_breakdown": {
"relevance": reward_obj.relevance,
"structure": reward_obj.structure,
"depth": reward_obj.depth,
"professionalism": reward_obj.professionalism,
"total_score": reward_obj.total_score,
},
"feedback": reward_obj.feedback,
"truncated": truncated,
"question_number": self._current_index, # 1-indexed (just answered)
"total_questions": len(self._task.questions),
}
if self._done:
final_score = self._final_score()
info["episode_summary"] = {
"final_score": final_score,
"per_question_scores": self._per_q_scores,
"passed": final_score >= 0.60,
"pass_threshold": 0.60,
}
return StepResult(
observation = obs,
reward = round(step_reward, 4),
done = self._done,
info = info,
)
def state(self) -> InterviewState:
"""
Return a serialisable snapshot of the full internal state.
Useful for debugging, checkpointing, and OpenEnv validation.
"""
return InterviewState(
task_id = self._task_id,
difficulty = self._task.difficulty,
job_title = self._task.job_title,
company_context = self._task.company_context,
questions = [q.text for q in self._task.questions],
stages = [q.stage for q in self._task.questions],
current_index = self._current_index,
qa_history = self._qa_history,
per_question_scores = self._per_q_scores,
episode_done = self._done,
total_steps = self._total_steps,
start_time = self._start_time,
cumulative_reward = self._cumulative_reward,
)
# ──────────────────────────────────────────
# Properties
# ──────────────────────────────────────────
@property
def task_id(self) -> str:
return self._task_id
@property
def available_tasks(self) -> list[str]:
return self.VALID_TASKS
@property
def action_space_description(self) -> str:
return (
"InterviewAction(response: str) — "
"A free-text string (1–2000 chars) representing the candidate's spoken answer."
)
@property
def observation_space_description(self) -> str:
return (
"InterviewObservation with fields: "
"task_id, job_title, company_context, difficulty, "
"current_question, question_number, total_questions, stage, "
"previous_qa (list of QAPair), cumulative_score, time_elapsed, interview_complete."
)
# ──────────────────────────────────────────
# Private helpers
# ──────────────────────────────────────────
def _build_observation(self) -> InterviewObservation:
if self._done or self._current_index >= len(self._task.questions):
current_question = (
"Thank you for completing the interview. "
"We will be in touch shortly."
)
stage = InterviewStage.COMPLETE
question_number = len(self._task.questions)
interview_complete = True
else:
q = self._task.questions[self._current_index]
current_question = q.text
stage = q.stage
question_number = self._current_index + 1
interview_complete = False
return InterviewObservation(
task_id = self._task_id,
job_title = self._task.job_title,
company_context = self._task.company_context,
difficulty = self._task.difficulty,
current_question = current_question,
question_number = question_number,
total_questions = len(self._task.questions),
stage = stage,
previous_qa = list(self._qa_history),
cumulative_score = round(self._cumulative_reward, 4),
time_elapsed = round(time.time() - self._start_time, 2),
interview_complete = interview_complete,
)
def _final_score(self) -> float:
"""Weighted average over all per-question scores."""
if not self._per_q_scores:
return 0.0
questions = self._task.questions
weighted = sum(
s * q.weight
for s, q in zip(self._per_q_scores, questions)
)
total_w = sum(q.weight for q in questions)
return round(weighted / total_w, 4) if total_w > 0 else 0.0
|