app.py

from transformers import FlaxAutoModelForSeq2SeqLM, AutoTokenizer, AutoModel
import torch
import numpy as np
import random
import json
from fastapi import FastAPI
from fastapi.responses import JSONResponse
from pydantic import BaseModel
# Keine Gradio-Imports hier!

# Lade RecipeBERT Modell (für semantische Zutat-Kombination)
bert_model_name = "alexdseo/RecipeBERT"
bert_tokenizer = AutoTokenizer.from_pretrained(bert_model_name)
bert_model = AutoModel.from_pretrained(bert_model_name)
bert_model.eval()

# Lade T5 Rezeptgenerierungsmodell
MODEL_NAME_OR_PATH = "flax-community/t5-recipe-generation"
t5_tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME_OR_PATH, use_fast=True)
t5_model = FlaxAutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME_OR_PATH)

# Token Mapping (bleibt gleich)
special_tokens = t5_tokenizer.all_special_tokens
tokens_map = {
    "<sep>": "--",
    "<section>": "\n"
}

# Deine Helper-Funktionen (get_embedding, average_embedding, get_cosine_similarity, etc.)
# ... diese bleiben ALLE GLEICH wie in deinem aktuellen app.py Code ...
# Kopiere alle Funktionen von 'get_embedding' bis 'generate_recipe_with_t5' hierher.
# (Ich kürze sie hier aus Platzgründen, aber sie müssen vollständig eingefügt werden)

def get_embedding(text):
    inputs = bert_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        outputs = bert_model(**inputs)
    attention_mask = inputs['attention_mask']
    token_embeddings = outputs.last_hidden_state
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    sum_embeddings = torch.sum(token_embeddings * input_mask_expanded, 1)
    sum_mask = torch.clamp(input_mask_expanded.sum(1), min=1e-9)
    return (sum_embeddings / sum_mask).squeeze(0)

def average_embedding(embedding_list):
    tensors = torch.stack([emb for _, emb in embedding_list])
    return tensors.mean(dim=0)

def get_cosine_similarity(vec1, vec2):
    if torch.is_tensor(vec1): vec1 = vec1.detach().numpy()
    if torch.is_tensor(vec2): vec2 = vec2.detach().numpy()
    vec1 = vec1.flatten()
    vec2 = vec2.flatten()
    dot_product = np.dot(vec1, vec2)
    norm_a = np.linalg.norm(vec1)
    norm_b = np.linalg.norm(vec2)
    if norm_a == 0 or norm_b == 0: return 0
    return dot_product / (norm_a * norm_b)

def get_combined_scores(query_vector, embedding_list, all_good_embeddings, avg_weight=0.6):
    results = []
    for name, emb in embedding_list:
        avg_similarity = get_cosine_similarity(query_vector, emb)
        individual_similarities = [get_cosine_similarity(good_emb, emb) for _, good_emb in all_good_embeddings]
        avg_individual_similarity = sum(individual_similarities) / len(individual_similarities)
        combined_score = avg_weight * avg_similarity + (1 - avg_weight) * avg_individual_similarity
        results.append((name, emb, combined_score))
    results.sort(key=lambda x: x[2], reverse=True)
    return results

def find_best_ingredients(required_ingredients, available_ingredients, max_ingredients=6, avg_weight=0.6):
    required_ingredients = list(set(required_ingredients))
    available_ingredients = list(set([i for i in available_ingredients if i not in required_ingredients]))
    if not required_ingredients and available_ingredients:
        random_ingredient = random.choice(available_ingredients)
        required_ingredients = [random_ingredient]
        available_ingredients = [i for i in available_ingredients if i != random_ingredient]
    if not required_ingredients or len(required_ingredients) >= max_ingredients:
        return required_ingredients[:max_ingredients]
    if not available_ingredients:
        return required_ingredients
    embed_required = [(e, get_embedding(e)) for e in required_ingredients]
    embed_available = [(e, get_embedding(e)) for e in available_ingredients]
    num_to_add = min(max_ingredients - len(required_ingredients), len(available_ingredients))
    final_ingredients = embed_required.copy()
    for _ in range(num_to_add):
        avg = average_embedding(final_ingredients)
        candidates = get_combined_scores(avg, embed_available, final_ingredients, avg_weight)
        if not candidates: break
        best_name, best_embedding, _ = candidates[0]
        final_ingredients.append((best_name, best_embedding))
        embed_available = [item for item in embed_available if item[0] != best_name]
    return [name for name, _ in final_ingredients]

def skip_special_tokens(text, special_tokens):
    for token in special_tokens: text = text.replace(token, "")
    return text

def target_postprocessing(texts, special_tokens):
    if not isinstance(texts, list): texts = [texts]
    new_texts = []
    for text in texts:
        text = skip_special_tokens(text, special_tokens)
        for k, v in tokens_map.items(): text = text.replace(k, v)
        new_texts.append(text)
    return new_texts

def validate_recipe_ingredients(recipe_ingredients, expected_ingredients, tolerance=0):
    recipe_count = len([ing for ing in recipe_ingredients if ing and ing.strip()])
    expected_count = len(expected_ingredients)
    return abs(recipe_count - expected_count) == tolerance

def generate_recipe_with_t5(ingredients_list, max_retries=5):
    original_ingredients = ingredients_list.copy()
    for attempt in range(max_retries):
        try:
            if attempt > 0:
                current_ingredients = original_ingredients.copy()
                random.shuffle(current_ingredients)
            else:
                current_ingredients = ingredients_list
            ingredients_string = ", ".join(current_ingredients)
            prefix = "items: "
            generation_kwargs = {
                "max_length": 512, "min_length": 64, "do_sample": True,
                "top_k": 60, "top_p": 0.95
            }
            inputs = t5_tokenizer(
                prefix + ingredients_string, max_length=256, padding="max_length",
                truncation=True, return_tensors="jax"
            )
            output_ids = t5_model.generate(
                input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, **generation_kwargs
            )
            generated = output_ids.sequences
            generated_text = target_postprocessing(t5_tokenizer.batch_decode(generated, skip_special_tokens=False), special_tokens)[0]
            recipe = {}
            sections = generated_text.split("\n")
            for section in sections:
                section = section.strip()
                if section.startswith("title:"):
                    recipe["title"] = section.replace("title:", "").strip().capitalize()
                elif section.startswith("ingredients:"):
                    ingredients_text = section.replace("ingredients:", "").strip()
                    recipe["ingredients"] = [item.strip().capitalize() for item in ingredients_text.split("--") if item.strip()]
                elif section.startswith("directions:"):
                    directions_text = section.replace("directions:", "").strip()
                    recipe["directions"] = [step.strip().capitalize() for step in directions_text.split("--") if step.strip()]
            if "title" not in recipe:
                recipe["title"] = f"Rezept mit {', '.join(current_ingredients[:3])}"
            if "ingredients" not in recipe:
                recipe["ingredients"] = current_ingredients
            if "directions" not in recipe:
                recipe["directions"] = ["Keine Anweisungen generiert"]
            if validate_recipe_ingredients(recipe["ingredients"], original_ingredients):
                return recipe
            else:
                if attempt == max_retries - 1: return recipe
        except Exception as e:
            if attempt == max_retries - 1:
                return {
                    "title": f"Rezept mit {original_ingredients[0] if original_ingredients else 'Zutaten'}",
                    "ingredients": original_ingredients,
                    "directions": ["Fehler beim Generieren der Rezeptanweisungen"]
                }
    return {
        "title": f"Rezept mit {original_ingredients[0] if original_ingredients else 'Zutaten'}",
        "ingredients": original_ingredients,
        "directions": ["Fehler beim Generieren der Rezeptanweisungen"]
    }

def process_recipe_request_logic(required_ingredients, available_ingredients, max_ingredients, max_retries):
    """
    Kernlogik zur Verarbeitung einer Rezeptgenerierungsanfrage.
    """
    if not required_ingredients and not available_ingredients:
        return {"error": "Keine Zutaten angegeben"}
    try:
        optimized_ingredients = find_best_ingredients(
            required_ingredients, available_ingredients, max_ingredients
        )
        recipe = generate_recipe_with_t5(optimized_ingredients, max_retries)
        result = {
            'title': recipe['title'],
            'ingredients': recipe['ingredients'],
            'directions': recipe['directions'],
            'used_ingredients': optimized_ingredients
        }
        return result
    except Exception as e:
        return {"error": f"Fehler bei der Rezeptgenerierung: {str(e)}"}

# --- FastAPI-Implementierung ---
app = FastAPI()

class RecipeRequest(BaseModel):
    required_ingredients: list[str] = []
    available_ingredients: list[str] = []
    max_ingredients: int = 7
    max_retries: int = 5
    # Optional: Für Abwärtskompatibilität, falls 'ingredients' als Top-Level-Feld gesendet wird
    # ingredients: list[str] = [] # Dies würde auch akzeptiert und müsste dann in der Logik verarbeitet werden

@app.post("/generate_recipe") # Einfacher Endpunkt, den Flutter aufruft
async def generate_recipe_api(request_data: RecipeRequest):
    """
    Standard-REST-API-Endpunkt für die Flutter-App.
    Nimmt direkt JSON-Daten an und gibt direkt JSON zurück.
    """
    # Verarbeite optionale Abwärtskompatibilität hier, falls nötig
    if not request_data.required_ingredients and 'ingredients' in request_data.model_dump():
        request_data.required_ingredients = request_data.model_dump()['ingredients']

    result_dict = process_recipe_request_logic(
        request_data.required_ingredients,
        request_data.available_ingredients,
        request_data.max_ingredients,
        request_data.max_retries
    )
    return JSONResponse(content=result_dict)

# In diesem Setup gibt es keine Gradio UI, nur die FastAPI-API.
# Dadurch sollte der Space zuverlässiger starten.
print("INFO: FastAPI application script finished execution and defined 'app' variable.")

# Der if __name__ == "__main__": Block wird von Hugging Face Spaces ignoriert,
# da sie den Uvicorn-Server direkt starten, der die 'app'-Variable sucht.