File size: 13,189 Bytes
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
500e2b8
d23cbe2
7743c15
 
 
 
 
500e2b8
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d23cbe2
7743c15
d23cbe2
500e2b8
d23cbe2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7743c15
 
 
 
 
 
 
500e2b8
 
 
 
 
 
 
 
 
 
 
7743c15
 
 
 
 
 
 
500e2b8
 
 
 
7743c15
 
013e24a
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d23cbe2
7743c15
 
 
 
 
 
 
 
 
 
 
 
4535620
d23cbe2
 
6475c9b
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d23cbe2
 
 
7743c15
 
d23cbe2
 
 
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d23cbe2
 
 
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d23cbe2
 
 
 
 
 
 
 
 
 
7743c15
d23cbe2
7743c15
 
d23cbe2
7743c15
 
d23cbe2
 
7743c15
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
ContentGuardEnv — Policy Enforcement Gateway (v1.0)
==================================================
This server acts as the primary interface between LLM agents and the 
ContentGuard moderation environment. It exposes a standardized OpenEnv 
OpenAPI/WebSocket interface to facilitate autonomous training and 
benchmarking across Meta community standards.

Key Features:
- Real-time policy-trace streaming via WebSockets.
- Dynamic environment resetting for multi-task RLHF.
- Automated grading & reward calculation.
"""

import os
import json
import time
from typing import Dict, Any, Optional
from dotenv import load_dotenv

load_dotenv()

import uvicorn
from fastapi import FastAPI, WebSocket, WebSocketDisconnect, HTTPException, Request
from fastapi.middleware.cors import CORSMiddleware
from fastapi.staticfiles import StaticFiles
from fastapi.responses import HTMLResponse
from pydantic import BaseModel, Field

from openai import AsyncOpenAI

# Internal module imports (Post-Restructure)
from env import ContentGuardEnv

app = FastAPI(
    title="ContentGuardEnv Gateway",
    description="Operational environment for Meta-scale Trust & Safety agent benchmarking.",
    version="1.0.0",
)

app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])

# Resource Path Configuration
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
STATIC_DIR = os.path.join(BASE_DIR, "static")

if os.path.exists(STATIC_DIR):
    app.mount("/static", StaticFiles(directory=STATIC_DIR), name="static")

sessions: Dict[str, ContentGuardEnv] = {}

# LLM Inference Client (Defaulting to Hackathon standard endpoints)
DEFAULT_BASE_URL = os.environ.get("API_BASE_URL", "https://api.openai.com/v1")
MODEL_NAME = os.environ.get("MODEL_NAME", "gpt-4o-mini")
DEFAULT_API_KEY = os.environ.get("HF_TOKEN") or os.environ.get("OPENAI_API_KEY")
BUILD_TAG = os.environ.get("HF_SPACE_SHA") or os.environ.get("SPACE_BUILD_TAG") or os.environ.get("BUILD_TAG") or str(int(time.time()))


def _is_placeholder_api_key(api_key: Optional[str]) -> bool:
    if not api_key:
        return True
    lowered = api_key.strip().lower()
    return lowered in {"sk-placeholder", "your_api_key", "changeme"}


def _resolve_session_client(cfg: Optional[Dict[str, Any]]) -> tuple[Optional[AsyncOpenAI], str]:
    """Build a runtime client from UI config with provider-aware routing rules."""
    if not cfg:
        return aclient, MODEL_NAME

    api_key = (cfg.get("api_key") or "").strip()
    base_url = (cfg.get("base_url") or "").strip() or DEFAULT_BASE_URL
    model = (cfg.get("model") or "").strip() or MODEL_NAME

    if _is_placeholder_api_key(api_key):
        # No runtime key provided: use server default if configured, otherwise deterministic grading fallback.
        return aclient, MODEL_NAME

    if api_key.startswith("hf_") and "openai.com" in base_url:
        base_url = "https://api-inference.huggingface.co/v1"
    elif api_key.startswith("sk-") and "huggingface.co" in base_url:
        base_url = "https://api.openai.com/v1"

    if api_key.startswith("hf_") and not (cfg.get("model") or "").strip():
        model = "meta-llama/Llama-3-70b-instruct"

    return AsyncOpenAI(api_key=api_key, base_url=base_url), model


def _build_demo_action(env: ContentGuardEnv) -> Dict[str, Any]:
    """Generate a task-valid deterministic action when live inference is unavailable."""
    gt = env.ground_truth or {}
    if env.task_id == "easy":
        return {"violation": gt.get("violation", "safe")}
    if env.task_id == "medium":
        return {
            "action": gt.get("action", "no_action"),
            "severity": int(gt.get("severity", 3)),
            "reasoning": "Deterministic demo fallback due unavailable inference credentials.",
        }
    return {
        "ruling": gt.get("ruling", "upheld"),
        "policy_references": gt.get("policy_references", []),
        "explanation": "Deterministic fallback path used because model inference is unavailable.",
        "user_guidance": "Review platform standards and avoid repeating flagged behavior.",
    }


aclient: Optional[AsyncOpenAI] = None
if not _is_placeholder_api_key(DEFAULT_API_KEY):
    aclient = AsyncOpenAI(api_key=DEFAULT_API_KEY.strip(), base_url=DEFAULT_BASE_URL)

class ResetRequest(BaseModel):
    task_id: str = Field(default="easy", description="Difficulty tier: easy | medium | hard")

class StepRequest(BaseModel):
    action: Dict[str, Any] = Field(..., description="Agent moderation decision package")


@app.middleware("http")
async def disable_cache_for_dashboard_assets(request: Request, call_next):
    response = await call_next(request)
    path = request.url.path
    if path == "/" or path.startswith("/static/"):
        response.headers["Cache-Control"] = "no-store, no-cache, must-revalidate, max-age=0"
        response.headers["Pragma"] = "no-cache"
        response.headers["Expires"] = "0"
    return response

@app.get("/", response_class=HTMLResponse, include_in_schema=False)
async def serve_dashboard():
    """Serves the primary autonomous monitoring interface."""
    index_path = os.path.join(STATIC_DIR, "index.html")
    if not os.path.exists(index_path):
        return HTMLResponse("ContentGuard Dashboard: Static assets not found. Check /server/static deployment.")
    with open(index_path, "r", encoding="utf-8") as f:
        html = f.read()
    html = html.replace('/static/style.css', f'/static/style.css?v={BUILD_TAG}', 1)
    html = html.replace('/static/app.js', f'/static/app.js?v={BUILD_TAG}', 1)
    return HTMLResponse(content=html)

@app.post("/reset", tags=["Core API"])
async def reset_environment(req: ResetRequest = ResetRequest()):
    """Initializes a new moderation case and returns the observation state."""
    env = ContentGuardEnv()
    obs = env.reset(task_id=req.task_id)
    sessions[env.episode_id] = env
    return obs

@app.post("/step/{episode_id}", tags=["Core API"])
async def process_step(episode_id: str, req: StepRequest):
    """Submits an agent decision and returns a policy-aligned reward signal."""
    env = sessions.get(episode_id)
    if not env:
        raise HTTPException(404, f"Session '{episode_id}' not active or expired.")
    try:
        return await env.step(req.action, client=aclient, model=MODEL_NAME)
    except RuntimeError as e:
        raise HTTPException(400, f"Policy Engine Conflict: {str(e)}")

@app.get("/state/{episode_id}", tags=["Advanced Utility"])
async def get_env_state(episode_id: str):
    """Retrieves the full internal state of an active moderation episode."""
    env = sessions.get(episode_id)
    if not env:
        raise HTTPException(404, f"Session '{episode_id}' not found.")
    return env.state()

@app.get("/health", tags=["System"])
async def check_health():
    return {"status": "operational", "active_sessions": len(sessions), "engine": "ContentGuardEnv"}

@app.websocket("/ws")
async def policy_trace_socket(websocket: WebSocket):
    """Streams real-time reasoning traces and environment telemetry."""
    await websocket.accept()
    env: ContentGuardEnv | None = None
    session_client: AsyncOpenAI | None = aclient
    session_model: str = MODEL_NAME
    
    try:
        while True:
            raw = await websocket.receive_text()
            try:
                msg = json.loads(raw)
            except json.JSONDecodeError:
                await websocket.send_json({"type": "error", "message": "Malformed WebSocket packet."})
                continue

            cmd = msg.get("action")
            # Universal Credential Injector (Session-based)
            if "config" in msg:
                session_client, session_model = _resolve_session_client(msg.get("config"))

            if cmd == "reset":
                env = ContentGuardEnv()
                try:
                    obs = env.reset(task_id=msg.get("task_id", "easy"))
                    # Explicitly dump Pydantic model for WebSocket JSON serialization
                    await websocket.send_json({"type": "reset", "observation": obs.model_dump()})
                except ValueError as e:
                    await websocket.send_json({"type": "error", "message": str(e)})
                    
            elif cmd == "step":
                if not env:
                    await websocket.send_json({"type": "error", "message": "State conflict: Submit reset before step."})
                    continue
                try:
                    result = await env.step(msg.get("data", {}), client=session_client, model=session_model)
                    await websocket.send_json({"type": "step", "result": result})
                except RuntimeError as e:
                    await websocket.send_json({"type": "error", "message": f"Execution halted: {str(e)}"})

            elif cmd == "run_agent":
                if not env:
                    await websocket.send_json({"type": "error", "message": "Session inactive."})
                    continue
                if env.done:
                    await websocket.send_json({"type": "error", "message": "Episode finished. Call reset() for a new case."})
                    continue
                try:
                    await websocket.send_json({"type": "stream", "content": f"[START] ep={env.episode_id} task={env.task_id}\n"})

                    if session_client is None:
                        raise RuntimeError("No API credentials configured.")
                    
                    sys_prompt = "Expert Safety Moderator. Respond with JSON only. Strictly align with platform policies."
                    user_prompt = f"Policy Task: {env._task_config['description']}\n\nEvidence:\n{json.dumps(env.case)}\n\nSubmit ruling in JSON."
                    
                    stream = await session_client.chat.completions.create(
                        model=session_model,
                        messages=[{"role": "system", "content": sys_prompt}, {"role": "user", "content": user_prompt}],
                        temperature=0.0,
                        stream=True
                    )
                    
                    full_response = ""
                    async for chunk in stream:
                        content = chunk.choices[0].delta.content
                        if content:
                            full_response += content
                            await websocket.send_json({"type": "stream", "content": content})

                    if not full_response.strip():
                        raise ValueError("Model returned an empty response.")
                    
                    # Clean/Parse Output
                    js_str = full_response.strip()
                    if js_str.startswith("```"):
                        js_str = js_str.split("```")[1]
                        if js_str.startswith("json"): js_str = js_str[4:]
                    
                    action = json.loads(js_str.strip())
                    await websocket.send_json({"type": "stream", "content": f"\n\n[STEP] Policy Ingested: {json.dumps(action)}\n"})
                    
                    result = await env.step(action, client=session_client, model=session_model)
                    await websocket.send_json({"type": "step", "result": result})
                    await websocket.send_json({"type": "stream", "content": f"[END] Result: Success. Reward: {result['reward']:.4f}\n"})
                    
                except Exception as e:
                    err_text = str(e)
                    lowered_err = err_text.lower()
                    if "invalid_api_key" in lowered_err or "incorrect api key" in lowered_err or "api key" in lowered_err or "401" in lowered_err:
                        err_text = "Authentication failed for the configured provider."

                    await websocket.send_json({"type": "stream", "content": f"\n\n[NOTICE] Inference Unavailable: {err_text}\nInitiating Passive Grader demo...\n"})

                    if env.done:
                        await websocket.send_json({"type": "error", "message": "Episode finished. Call reset() for a new case."})
                        continue
                    
                    sim_action = _build_demo_action(env)
                    
                    try:
                        result = await env.step(sim_action, client=None, model=session_model)
                        await websocket.send_json({"type": "step", "result": result})
                        await websocket.send_json({"type": "stream", "content": f"\n[DEMO] Passive Ruling Emitted. Final Reward: {result['reward']:.4f}\n"})
                    except RuntimeError as step_error:
                        await websocket.send_json({"type": "error", "message": str(step_error)})
            
            elif cmd == "state":
                if env: await websocket.send_json({"type": "state", "state": env.state()})
    except WebSocketDisconnect:
        pass

def main():
    """Server entry point for OpenEnv validation."""
    print("🚀 Initializing ContentGuard Policy Gateway...")
    uvicorn.run("app:app", host="0.0.0.0", port=7860, reload=False)

if __name__ == "__main__":
    main()