File size: 11,378 Bytes
743203e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
"""
Inference Script for the Policy-to-Logic RL Environment
=======================================================

MANDATORY ENVIRONMENT VARIABLES:
    API_BASE_URL   The API endpoint for the LLM.
    MODEL_NAME     The model identifier to use for inference.
    HF_TOKEN       Your Hugging Face / API key.

STDOUT FORMAT:
    [START] task=<task_name> env=<benchmark> model=<model_name>
    [STEP]  step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
    [END]   success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...,rn>

This script runs the LLM agent against all 3 tasks in the environment
and produces reproducible baseline scores.
"""

import json
import os
import sys
import textwrap
from typing import List, Optional

from openai import OpenAI

# ─── Add parent to path for imports ──────────────────────────────
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

from policy_to_logic_env.client import PolicyToLogicEnv
from policy_to_logic_env.models import PolicyToLogicAction

# ─── Configuration ───────────────────────────────────────────────
API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
ENV_BASE_URL = os.getenv("ENV_BASE_URL", "http://localhost:7860")
BENCHMARK = "policy_to_logic"

TASKS = ["data_access", "resource_access", "transaction_approval"]
TEMPERATURE = 0.3
MAX_TOKENS = 1024


# ─── Logging ─────────────────────────────────────────────────────
def log_start(task: str, env: str, model: str) -> None:
    print(f"[START] task={task} env={env} model={model}", flush=True)


def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
    error_val = error if error else "null"
    done_val = str(done).lower()
    # Truncate action for readability
    action_short = action[:80].replace("\n", " ") if action else "none"
    print(
        f"[STEP] step={step} action={action_short} reward={reward:.2f} done={done_val} error={error_val}",
        flush=True,
    )


def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
    rewards_str = ",".join(f"{r:.2f}" for r in rewards)
    print(
        f"[END] success={str(success).lower()} steps={steps} score={score:.2f} rewards={rewards_str}",
        flush=True,
    )


# ─── System Prompt ───────────────────────────────────────────────
SYSTEM_PROMPT = textwrap.dedent("""
You are an AI agent interacting with a Policy-to-Logic RL environment.

Your goal is to convert natural language access control policies into
executable logic rules using a simple JSON-based DSL.

AVAILABLE ACTIONS:
1. ask_clarification - Ask a question about unclear policy terms
2. propose_rules - Submit a complete rule set
3. refine_rules - Improve your previous rules based on feedback

RESPONSE FORMAT:
You must respond with a valid JSON object:

For asking clarification:
{"action_type": "ask_clarification", "content": "{\\"question\\": \\"What are business hours?\\"}"}

For proposing or refining rules:
{"action_type": "propose_rules", "content": "{\\"rules\\": [{\\"if\\": [{\\"field\\": \\"time\\", \\"op\\": \\">\\", \\"value\\": 18}, {\\"field\\": \\"data_type\\", \\"op\\": \\"==\\", \\"value\\": \\"sensitive\\"}], \\"then\\": \\"DENY\\"}], \\"default\\": \\"ALLOW\\"}"}

DSL FORMAT:
- Rules are a list of IF-THEN objects
- All conditions in a rule are AND-ed
- Rules evaluated top-to-bottom, first match wins
- Operators: >, <, >=, <=, ==, !=

STRATEGY:
- Read the policy carefully
- Ask 1-2 clarification questions for ambiguous terms (thresholds, hours)
- Propose rules covering all cases
- Refine based on test result feedback
- Aim for >90% accuracy to complete the episode

Respond with ONLY the JSON object, nothing else.
""").strip()


# ─── Agent Logic ─────────────────────────────────────────────────
def build_user_prompt(observation: dict, step: int, history: List[str]) -> str:
    """Build the user prompt from the current observation."""
    parts = [f"Step {step} of {observation.get('max_steps', 7)}"]

    parts.append(f"\n--- POLICY ---\n{observation.get('policy_text', '')}")

    if observation.get("feedback"):
        parts.append(f"\n--- FEEDBACK ---\n{observation['feedback']}")

    if observation.get("clarification_response"):
        parts.append(f"\n--- CLARIFICATION ANSWER ---\n{observation['clarification_response']}")

    if observation.get("test_results"):
        tr = observation["test_results"]
        parts.append(
            f"\n--- TEST RESULTS ---\n"
            f"Passed: {tr.get('passed', 0)}/{tr.get('total', 0)}\n"
            f"Accuracy: {observation.get('current_accuracy', 0):.1%}"
        )
        if tr.get("sample_failures"):
            parts.append("Sample failures:")
            for fail in tr["sample_failures"][:3]:
                parts.append(f"  Scenario: {fail.get('scenario', {})}")
                parts.append(f"  Expected: {fail.get('expected', '?')}, Got: {fail.get('got', '?')}")

    parts.append(f"\n--- DSL FORMAT ---\n{observation.get('dsl_format', '')}")
    parts.append(f"\nAvailable actions: {observation.get('available_actions', [])}")
    parts.append(f"Current accuracy: {observation.get('current_accuracy', 0):.1%}")

    if history:
        parts.append(f"\n--- HISTORY ---")
        for h in history[-3:]:
            parts.append(h)

    parts.append("\nRespond with your next action as a JSON object.")

    return "\n".join(parts)


def get_agent_action(
    client: OpenAI, observation: dict, step: int, history: List[str]
) -> tuple[str, str]:
    """
    Query the LLM to get the next action.

    Returns (action_type, content)
    """
    user_prompt = build_user_prompt(observation, step, history)

    try:
        completion = client.chat.completions.create(
            model=MODEL_NAME,
            messages=[
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
            temperature=TEMPERATURE,
            max_tokens=MAX_TOKENS,
            stream=False,
        )
        raw = (completion.choices[0].message.content or "").strip()

        # Parse the response
        # Try to extract JSON from the response
        try:
            # Handle markdown code blocks
            if "```json" in raw:
                raw = raw.split("```json")[1].split("```")[0].strip()
            elif "```" in raw:
                raw = raw.split("```")[1].split("```")[0].strip()

            parsed = json.loads(raw)
            action_type = parsed.get("action_type", "propose_rules")
            content = parsed.get("content", "{}")

            # If content is already a dict, serialize it
            if isinstance(content, dict):
                content = json.dumps(content)

            return action_type, content

        except (json.JSONDecodeError, IndexError):
            # Fallback: try to extract rules from raw text
            if '"rules"' in raw:
                return "propose_rules", raw
            else:
                return "propose_rules", json.dumps({
                    "rules": [],
                    "default": "ALLOW"
                })

    except Exception as exc:
        print(f"[DEBUG] Model request failed: {exc}", flush=True)
        return "propose_rules", json.dumps({"rules": [], "default": "ALLOW"})


# ─── Main Loop ───────────────────────────────────────────────────
def run_task(llm_client: OpenAI, env_client: PolicyToLogicEnv, task_name: str) -> float:
    """Run a single task and return the score."""
    history: List[str] = []
    rewards: List[float] = []
    steps_taken = 0
    score = 0.0
    success = False

    log_start(task=task_name, env=BENCHMARK, model=MODEL_NAME)

    try:
        # Reset environment
        result = env_client.reset(task_name=task_name)
        obs = result.observation.model_dump()
        max_steps = obs.get("max_steps", 7)

        for step in range(1, max_steps + 1):
            if result.done:
                break

            # Get agent action
            action_type, content = get_agent_action(llm_client, obs, step, history)

            # Step the environment
            action = PolicyToLogicAction(action_type=action_type, content=content)
            result = env_client.step(action)
            obs = result.observation.model_dump()

            reward = result.reward
            done = result.done
            error = result.info.get("error") if result.info else None

            rewards.append(reward)
            steps_taken = step

            log_step(
                step=step,
                action=f"{action_type}({content[:50]}...)" if len(content) > 50 else f"{action_type}({content})",
                reward=reward,
                done=done,
                error=error,
            )

            history.append(
                f"Step {step}: {action_type} β†’ accuracy={obs.get('current_accuracy', 0):.1%}, reward={reward:.2f}"
            )

            if done:
                break

        # Compute final score
        final_accuracy = obs.get("current_accuracy", 0.0)
        score = final_accuracy  # Clean score = accuracy
        score = min(max(score, 0.0), 1.0)
        success = score >= 0.5

    except Exception as e:
        print(f"[DEBUG] Task {task_name} error: {e}", flush=True)

    finally:
        log_end(success=success, steps=steps_taken, score=score, rewards=rewards)

    return score


def main() -> None:
    """Run inference against all tasks."""
    if not API_KEY:
        print("[ERROR] HF_TOKEN or API_KEY not set. Please set one of these environment variables.", flush=True)
        sys.exit(1)

    # Initialize clients
    llm_client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
    env_client = PolicyToLogicEnv(base_url=ENV_BASE_URL)

    print(f"[INFO] Running inference with model={MODEL_NAME}", flush=True)
    print(f"[INFO] Environment: {ENV_BASE_URL}", flush=True)
    print(f"[INFO] Tasks: {TASKS}", flush=True)
    print("", flush=True)

    scores = {}

    for task_name in TASKS:
        print(f"\n{'='*60}", flush=True)
        print(f"Running task: {task_name}", flush=True)
        print(f"{'='*60}", flush=True)

        score = run_task(llm_client, env_client, task_name)
        scores[task_name] = score

    # Summary
    print(f"\n{'='*60}", flush=True)
    print("SUMMARY", flush=True)
    print(f"{'='*60}", flush=True)
    for task, score in scores.items():
        print(f"  {task}: {score:.2f}", flush=True)
    avg = sum(scores.values()) / len(scores) if scores else 0.0
    print(f"  Average: {avg:.2f}", flush=True)

    env_client.close()


if __name__ == "__main__":
    main()