engine_cloud.py

"""
Cloud inference connector for Pocket Dungeon Master.

Streams RPG narrative from the Modal-hosted vLLM backend (OpenAI-compatible).
Configured via environment variables (set as HF Space secrets):
    MODAL_BACKEND_URL  e.g. https://<workspace>--pocket-dm-backend-serve.modal.run
    MODAL_API_KEY      the VLLM_API_KEY chosen when deploying the Modal secret
"""
import os
from openai import OpenAI

from engine import load_system_prompt, log_trace

MODAL_BACKEND_URL = os.environ.get(
    "MODAL_BACKEND_URL", "https://your-unique-modal-subdomain.modal.run"
)
MODAL_API_KEY = os.environ.get("MODAL_API_KEY", "EMPTY")
MODEL_NAME = os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-3B-Instruct")  # tiny ≤4B (Tiny Titan)

_client = OpenAI(
    base_url=MODAL_BACKEND_URL.rstrip("/") + "/v1",
    api_key=MODAL_API_KEY or "EMPTY",
    timeout=120.0,
)


def call_modal_generator(item_id, history_messages):
    """Yields the cumulative narrative text as it streams from the backend."""
    messages = [{"role": "system", "content": load_system_prompt(item_id)}]
    for msg in history_messages:
        role = getattr(msg, "role", None) or (msg.get("role") if isinstance(msg, dict) else None)
        content = getattr(msg, "content", None) or (msg.get("content") if isinstance(msg, dict) else None)
        if content:
            messages.append({"role": role, "content": content})

    full_reply_text = ""
    try:
        stream = _client.chat.completions.create(
            model=MODEL_NAME,
            messages=messages,
            max_tokens=220,
            temperature=0.85,
            stream=True,
        )
        for chunk in stream:
            delta = chunk.choices[0].delta.content or ""
            if delta:
                full_reply_text += delta
                yield full_reply_text
    except Exception as e:
        yield f"⚠️ The artifact's voice falters — the cloud core is unreachable ({e})."
        return

    # Log the interaction (Sharing is Caring badge / public trace).
    try:
        last_user = next(
            (m["content"] for m in reversed(messages) if m["role"] == "user"), ""
        )
        log_trace({"item": item_id, "user": last_user, "model": full_reply_text})
    except Exception:
        pass


def get_combat_keywords():
    return ['rage', 'boil', 'heat', 'kill', 'attack', 'shatter', 'fight',
            'force', 'mutate', 'scream', 'wield', 'blade', 'potion', 'key']