models/llm.py

import json
from transformers import pipeline

class StyleSavvy:
    def __init__(
        self,
        model_name: str = "google/flan-t5-large",
        device: int   = -1,     # -1 = CPU, or GPU index
        max_length: int = 150,
    ):
        self.pipe = pipeline(
            "text2text-generation",
            model=model_name,
            tokenizer=model_name,
            device=device,
        )
        self.max_length = max_length

    def advise(self, items, body_type, face_shape, occasion):
        # 如果 items 是 JSON 字串，先 parse 回 Python list
        items = json.loads(items) if isinstance(items, str) else items

        prompt = (
            f"The user is {body_type}-shaped with a {face_shape} face, "
            f"attending a {occasion}. They are wearing: "
            + ", ".join(i["label"] for i in items)
            + ".\n\nPlease list 5 concise style tips as bullet points:"
        )

        result = self.pipe(
            prompt,
            max_length=self.max_length,
            num_beams=4,
            early_stopping=True,
            do_sample=False
        )[0]["generated_text"].strip()

        return result





# import torch

# # models/llm.py

# # models/llm.py

# import os
# from typing import List
# from transformers import pipeline, Pipeline

# # Force CPU modes (avoid any MPS/CUDA issues on macOS)
# os.environ["CUDA_VISIBLE_DEVICES"] = ""
# os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "0"

# class StyleSavvy:
#     def __init__(
#         self,
#         model_name: str = "openlm-research/open_llama_3b_v2",
#         device: int   = -1,        # -1 = CPU
#         max_new_tokens: int = 100,
#         temperature: float   = 0.7,
#         top_p: float         = 0.9,
#     ):
#         """
#         Uses OpenLLaMA-3B-v2 (≈3B params) for fast, local inference.
#         """
#         # Setup a causal text-generation pipeline
#         self.pipe: Pipeline = pipeline(
#             "text-generation",
#             model=model_name,
#             tokenizer=model_name,
#             device=device,
#         )
#         # GPT‐style models need a pad token to avoid warnings
#         if self.pipe.tokenizer.pad_token_id is None:
#             self.pipe.tokenizer.pad_token = self.pipe.tokenizer.eos_token
        
#         self.max_new_tokens = max_new_tokens
#         self.temperature    = temperature
#         self.top_p          = top_p

#     def advise(
#         self,
#         items: List[str],
#         body_type: str,
#         face_shape: str,
#         occasion: str
#     ) -> List[str]:
#         """
#         Builds a strict instruction prompt and returns exactly five "- " bullets.
#         """
#         labels = ", ".join(items) if items else "an outfit"
#         prompt = (
#             "You are a professional fashion consultant.\n"
#             f"The user is {body_type}-shaped with a {face_shape} face, attending {occasion}.\n"
#             f"They are wearing: {labels}.\n\n"
#             "Please provide exactly five concise style tips, each on its own line, "
#             "and starting with \"- \". No extra text."
#         )

#         # Generate
#         output = self.pipe(
#             prompt,
#             max_new_tokens=self.max_new_tokens,
#             do_sample=True,
#             temperature=self.temperature,
#             top_p=self.top_p,
#             return_full_text=False,
#         )[0]["generated_text"]

#         # Extract bullets
#         tips = [ln.strip() for ln in output.splitlines() if ln.strip().startswith("- ")]
#         # Fallback: split on sentences if fewer than 5 bullets
#         if len(tips) < 5:
#             candidates = [s.strip() for s in output.replace("\n"," ").split(".") if s.strip()]
#             tips = [f"- {candidates[i]}" for i in range(min(5, len(candidates)))]
        
#         return tips[:5]