"""Vision client — asks the model to guess what a doodle depicts, and draws SVG pictures for the player to guess. guess(pil_image) -> list[str] of lowercase one/two-word guesses (best first). draw_svg(word) -> str of SVG markup (or "" on failure). """ from __future__ import annotations import base64 import io import random import re import sys import requests import config _resolved_model = None # cached result of model-name auto-detection _warned = set() # so we only print each backend warning once def _warn_once(msg: str): if msg not in _warned: _warned.add(msg) print(f"[vision_client] {msg}", file=sys.stderr) def _auth_headers(): h = {"Content-Type": "application/json"} if config.MODEL_API_KEY: h["Authorization"] = f"Bearer {config.MODEL_API_KEY}" return h def _model_name(): """The id to send in the `model` field. If MODEL_NAME is 'auto' (or empty), ask the server's /v1/models and use what it actually serves — llama-server ignores this field, but vLLM validates it, so auto-detection lets one config drive either backend (and the llama.cpp gguf name need not be known).""" global _resolved_model name = (config.MODEL_NAME or "").strip() if name and name.lower() != "auto": return name if _resolved_model: return _resolved_model try: r = requests.get(f"{config.MODEL_BASE_URL}/v1/models", headers=_auth_headers(), timeout=10) r.raise_for_status() _resolved_model = r.json()["data"][0]["id"] except Exception: _resolved_model = "default" # llama-server accepts any id return _resolved_model def health(): """(ok, detail) for the configured backend — used for a startup check.""" if config.MOCK_MODE: return True, "MOCK mode (no model server set)" base = config.MODEL_BASE_URL for path in ("/health", "/v1/models"): # llama.cpp has /health; vLLM has /v1/models try: r = requests.get(f"{base}{path}", headers=_auth_headers(), timeout=5) if r.ok: return True, f"{base} — model '{_model_name()}'" except Exception as e: last = e.__class__.__name__ return False, f"{base} unreachable ({locals().get('last', 'no response')})" _GUESS_INTRO = ( "We are playing Pictionary. This is a rough, simple doodle drawn by a human. " "Guess what or WHO it depicts — it may be an object, a brand/logo, a character, " "or a famous person (use clues like jersey numbers, initials, symbols). Be " "SPECIFIC: if it is a soccer ball say 'soccer ball' not 'ball'; if the clues " "point to a person, name them." ) _GUESS_FORMAT = ( "You are told the answer's letter count, number of words, and a dash pattern " "(revealed letters are filled in) — every guess MUST have exactly that many " "letters and words and match the revealed letters.\n" "Respond in EXACTLY two lines, nothing else:\n" "THINK: one short sentence describing the shapes you see and your reasoning — REQUIRED, always include this even if uncertain.\n" "GUESS: your top 3 guesses, most likely first, comma-separated, lowercase." ) _DRAW_SVG_PROMPT = """Draw "{word}" as a Pictionary SVG that players can CLEARLY recognize and guess. Think step by step: 1. What is the most iconic silhouette or shape of "{word}"? 2. What 2-3 key details make it instantly recognizable? 3. What realistic colors should I use? Then draw it. SVG rules: - viewBox="0 0 500 360" width="500" height="360" - The canvas background is already WHITE — do NOT add a background rect and do NOT set fill on the