server/app.py

"""
HF Spaces server - exposes the AttentionEnv via a REST API
compatible with the OpenEnv spec (reset / step / state endpoints).
"""

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from typing import Dict, List, Optional
import uvicorn

from env.environment import AttentionEnv
from env.models import Action
from env.tasks import task_easy, task_medium, task_hard, grade_easy, grade_medium, grade_hard

app = FastAPI(title="Attention Allocation System", version="1.0.0")

_env: Optional[AttentionEnv] = None


def get_env() -> AttentionEnv:
    global _env
    if _env is None:
        _env = AttentionEnv()
    return _env


# ── Request models ─────────────────────────────────────────────────────────────
class StepRequest(BaseModel):
    item_id: int

class BaselineResponse(BaseModel):
    scores: Dict[str, float]


class GraderResponse(BaseModel):
    task: str
    score: float
    total_reward: float


class TasksResponse(BaseModel):
    tasks: List[str]
    action_schema: Dict[str, str]



# ── Endpoints ──────────────────────────────────────────────────────────────────
@app.get("/")
def root():
    return {
        "name": "Attention Allocation System",
        "description": "Content recommendation RL environment (OpenEnv compatible)",
        "endpoints": ["/reset", "/step", "/state", "/health", "/baseline", "/grader", "/tasks"],
    }


@app.get("/health")
def health():
    """Hackathon automated ping - must return 200."""
    return {"status": "ok"}


@app.post("/reset")
def reset():
    """Start a new episode. Returns the initial observation."""
    env = get_env()
    obs = env.reset()
    return {"observation": obs.model_dump(), "done": False}


@app.post("/step")
def step(request: StepRequest):
    """Take one action. Returns next observation, reward, and done flag."""
    env = get_env()
    try:
        action = Action(item_id=request.item_id)
        obs, reward, done, info = env.step(action)
        return {
            "observation": obs.model_dump(),
            "reward": reward.value,
            "done": done,
            "info": info,
        }
    except Exception as e:
        raise HTTPException(status_code=400, detail=str(e))


@app.get("/state")
def state():
    """Get current observation without taking an action."""
    env = get_env()
    if env.user is None:
        raise HTTPException(status_code=400, detail="Call /reset first")
    return env.state().model_dump()


@app.get("/baseline", response_model=BaselineResponse)
def baseline():
    """
    Run baseline inference on all 3 tasks (easy, medium, hard).
    Returns the baseline score for each task.
    """
    from openai import OpenAI
    import os

    api_key = os.getenv("OPENAI_API_KEY") or os.getenv("HF_TOKEN") or os.getenv("API_KEY")
    if not api_key:
        raise HTTPException(status_code=500, detail="No API key configured for baseline inference")
    
    api_base_url = os.getenv("API_BASE_URL", "https://api.groq.com/openai/v1")
    model_name = os.getenv("MODEL_NAME", "llama-3.1-8b-instant")
    
    client = OpenAI(base_url=api_base_url, api_key=api_key)
    
    from inference import run_episode
    
    scores = {}
    tasks = [
        ("easy", task_easy, 4.0),
        ("medium", task_medium, 7.0),
        ("hard", task_hard, 11.0),
    ]
    
    try:
        for task_name, task_fn, norm in tasks:
            env = task_fn()
            score = run_episode(client, env, task_name, norm)
            scores[task_name] = score
        
        return BaselineResponse(scores=scores)
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Baseline inference failed: {str(e)}")


@app.get("/grader")
def grader():
    """
    Calculate the grader score for the current episode.
    Returns the normalized score based on the number of items in the current environment.
    """
    env = get_env()
    if env.user is None:
        raise HTTPException(status_code=400, detail="Call /reset first")

    if not hasattr(env, 'history') or len(env.history) == 0:
        raise HTTPException(status_code=400, detail="No actions taken yet in the episode")
    
    total_reward = getattr(env, 'total_reward', 0.0)
    
    if env.num_items == 5:
        task_name = "easy"
        score = grade_easy(total_reward)
    elif env.num_items == 10:
        task_name = "medium"
        score = grade_medium(total_reward)
    elif env.num_items == 15:
        task_name = "hard"
        score = grade_hard(total_reward)
    else:
        raise HTTPException(status_code=400, detail=f"Unknown task with {env.num_items} items")
    
    return GraderResponse(task=task_name, score=score, total_reward=total_reward)


@app.get("/tasks", response_model=TasksResponse)
def tasks():
    """
    Returns the available tasks and the action schema required for each step.
    """
    return TasksResponse(
        tasks=["easy", "medium", "hard"],
        action_schema={
            "item_id": "integer (ID of the item to recommend)"
        }
    )



def main():
    uvicorn.run(app, host="0.0.0.0", port=7860)


if __name__ == "__main__":
    main()