app.py

import os
import json
import torch
import torch.nn as nn
from torch.nn import functional as F
from tokenizers import ByteLevelBPETokenizer
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from huggingface_hub import hf_hub_download
import urllib.request
import urllib.parse
import re

# --- 1. INITIALIZE FASTAPI API ---
app = FastAPI(title="Orbit SpaceStar Multi-Version Backend API", version="0.03_Super")

# Zorg dat je website (frontend) vlekkeloos mag praten met deze backend
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f"🛰️ Orbit Multi-Engine active on device: {DEVICE.upper()}")

# --- 2. DYNAMIC ARCHITECTURE CONFIGURATION ---
MODEL_CONFIGS = {
    "0.01": {"block_size": 256, "n_embd": 384, "n_head": 6, "n_layer": 8},
    "0.02": {"block_size": 256, "n_embd": 384, "n_head": 6, "n_layer": 8}
}
MODEL_REPOS = {
    "0.01": "Littendekitten/Orbit-SpaceStar-0.01",
    "0.02": "Littendekitten/Orbit-SpaceStar-0.02"
}
loaded_models = {}
loaded_tokenizers = {}

# --- 3. TRANSFORMER DYNAMIC ARCHITECTURE ---
class Head(nn.Module):
    def __init__(self, n_embd, head_size, block_size):
        super().__init__()
        self.key = nn.Linear(n_embd, head_size, bias=False)
        self.query = nn.Linear(n_embd, head_size, bias=False)
        self.value = nn.Linear(n_embd, head_size, bias=False)
        self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
        
    def forward(self, x):
        B, T, C = x.shape
        k, q = self.key(x), self.query(x)
        wei = q @ k.transpose(-2, -1) * k.shape[-1]**-0.5
        wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
        wei = F.softmax(wei, dim=-1)
        v = self.value(x)
        return wei @ v

class MultiHeadAttention(nn.Module):
    def __init__(self, n_embd, num_heads, head_size, block_size):
        super().__init__()
        self.heads = nn.ModuleList([Head(n_embd, head_size, block_size) for _ in range(num_heads)])
        self.proj = nn.Sequential(nn.Linear(head_size * num_heads, n_embd))
        
    def forward(self, x):
        out = torch.cat([h(x) for h in self.heads], dim=-1)
        return self.proj(out)

class FeedForward(nn.Module):
    def __init__(self, n_embd):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(n_embd, 4 * n_embd),
            nn.ReLU(),
            nn.Linear(4 * n_embd, n_embd),
        )
    def forward(self, x):
        return self.net(x)

class Block(nn.Module):
    def __init__(self, n_embd, n_head, block_size):
        super().__init__()
        head_size = n_embd // n_head
        self.sa = MultiHeadAttention(n_embd, n_head, head_size, block_size)
        self.ffwd = FeedForward(n_embd)
        self.ln1 = nn.LayerNorm(n_embd)
        self.ln2 = nn.LayerNorm(n_embd)
        
    def forward(self, x):
        x = x + self.sa(self.ln1(x))
        x = x + self.ffwd(self.ln2(x))
        return x

class OrbitTransformer(nn.Module):
    def __init__(self, vocab_size, config):
        super().__init__()
        self.block_size = config["block_size"]
        self.token_embedding_table = nn.Embedding(vocab_size, config["n_embd"])
        self.position_embedding_table = nn.Embedding(config["block_size"], config["n_embd"])
        self.blocks = nn.Sequential(*[Block(config["n_embd"], n_head=config["n_head"], block_size=config["block_size"]) for _ in range(config["n_layer"])])
        self.ln_f = nn.LayerNorm(config["n_embd"])
        self.lm_head = nn.Linear(config["n_embd"], vocab_size)
        
    def forward(self, idx):
        B, T = idx.shape
        tok_emb = self.token_embedding_table(idx)
        pos_emb = self.position_embedding_table(torch.arange(T, device=idx.device))
        x = tok_emb + pos_emb
        x = self.blocks(x)
        x = self.ln_f(x)
        logits = self.lm_head(x)
        return logits, None

    def generate(self, idx, max_new_tokens, tokenizer, temperature=0.5, top_k=30):
        """Geüpgradede generatie met een keiharde herhalingsrem (Repetition Penalty)"""
        end_token_id = tokenizer.token_to_id("<|end|>")
        repetition_penalty = 1.3  # Strafpunten voor tokens die hij al gebruikt heeft (voorkomt loops!)
        
        for _ in range(max_new_tokens):
            idx_cond = idx[:, -self.block_size:]
            logits, _ = self(idx_cond)
            logits = logits[:, -1, :] / temperature
            
            # REPETITION PENALTY: Als het model een token herhaalt, maken we de kans erop kleiner!
            for token_id in set(idx[0].tolist()):
                if logits[0, token_id] > 0:
                    logits[0, token_id] /= repetition_penalty
                else:
                    logits[0, token_id] *= repetition_penalty
            
            if top_k is not None:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = float('-inf')
                
            probs = F.softmax(logits, dim=-1)
            idx_next = torch.multinomial(probs, num_samples=1)
            
            idx = torch.cat((idx, idx_next), dim=1)
            
            # KEIHARDE STOP-TOKEN CHECK: Meteen kappen als hij klaar is!
            if end_token_id is not None and idx_next.item() == end_token_id:
                break
                
        return idx

# --- 4. DYNAMIC MODEL LOADER ---
def get_model_and_tokenizer(version: str):
    if version in loaded_models:
        return loaded_models[version], loaded_tokenizers[version]
    
    if version not in MODEL_REPOS:
        raise HTTPException(status_code=400, detail=f"Versie {version} niet ondersteund.")
    
    repo_id = MODEL_REPOS[version]
    config = MODEL_CONFIGS.get(version, MODEL_CONFIGS["0.02"])
    print(f"📥 Laden van versie {version} vanuit {repo_id}...")
    
    try:
        model_path = hf_hub_download(repo_id=repo_id, filename="best_model.pth")
        meta_path = hf_hub_download(repo_id=repo_id, filename="OrbitTokenizer/vocab_meta.json")
        vocab_path = hf_hub_download(repo_id=repo_id, filename="OrbitTokenizer/vocab.json")
        merges_path = hf_hub_download(repo_id=repo_id, filename="OrbitTokenizer/merges.txt")
        
        with open(meta_path, 'r', encoding='utf-8') as f:
            vocab_meta = json.load(f)
        vocab_size = vocab_meta["vocab_size"]
        
        tokenizer = ByteLevelBPETokenizer(vocab_path, merges_path, add_prefix_space=True)
        tokenizer.normalizer = None
        
        model = OrbitTransformer(vocab_size, config)
        model.load_state_dict(torch.load(model_path, map_location=DEVICE))
        model.to(DEVICE)
        model.eval()
        
        loaded_models[version] = model
        loaded_tokenizers[version] = tokenizer
        return model, tokenizer
        
    except Exception as e:
        if version == "0.02":
            print(f"⚠️ Proxy: Versie 0.02 faalt ({e}). Terugval naar 0.01!")
            m01, t01 = get_model_and_tokenizer("0.01")
            loaded_models["0.02"] = m01
            loaded_tokenizers["0.02"] = t01
            return m01, t01
        else:
            raise HTTPException(status_code=404, detail=f"Bestanden niet gevonden: {e}")

# --- 5. INTERNET SEARCH ENGINE (STRENGER AFGESTELD) ---
def fetch_internet_context(query: str) -> str:
    """Haalt info op via DuckDuckGo, maar ALLEEN bij duidelijke zoek-prompts."""
    q_lower = query.lower().strip()
    
    # Alleen zoeken als het echt een vraag is die begint met deze woorden
    triggers = ["zoek naar", "wat is", "wie is", "hoe werkt", "search for", "what is", "who is", "tell me about"]
    needs_internet = any(q_lower.startswith(t) for t in triggers)
    
    # Extra handrem: als de prompt super kort is (zoals "HI" of "ramen"), NOOIT internet gebruiken!
    if not needs_internet or len(q_lower) <= 4:
        return ""

    try:
        clean_q = q_lower
        for t in triggers:
            if clean_q.startswith(t):
                clean_q = clean_q[len(t):].strip()
                break
                
        search_term = urllib.parse.quote(clean_q if clean_q else query)
        url = f"https://api.duckduckgo.com/?q={search_term}&format=json&no_html=1"
        req = urllib.request.Request(url, headers={'User-Agent': 'Mozilla/5.0'})
        
        with urllib.request.urlopen(req, timeout=3) as response:
            data = json.loads(response.read().decode('utf-8'))
            abstract = data.get("AbstractText", "")
            if not abstract and data.get("RelatedTopics"):
                for item in data["RelatedTopics"]:
                    if isinstance(item, dict) and "Text" in item:
                        abstract = item["Text"]
                        break
            if abstract:
                return abstract[:180].strip()
    except Exception as e:
        print(f"📡 API fout: {e}")
    return ""

# --- 6. API ENDPOINTS ---
class ChatRequest(BaseModel):
    prompt: str
    version: str = "0.02"
    temperature: float = 0.5 # Iets lager gezet voor meer stabiliteit en minder hallucinaties
    max_tokens: int = 150    # Iets korter gezet zodat hij niet te lang door-rabbelt

@app.get("/")
def home():
    return {"status": "Orbit API is Online", "device": DEVICE}

@app.post("/chat")
def chat(request: ChatRequest):
    model, tokenizer = get_model_and_tokenizer(request.version)
    
    # Internet check (negeert "HI")
    internet_info = fetch_internet_context(request.prompt)
    
    think_html_block = ""
    
    # Bouw de nette chat-prompt
    if internet_info:
        think_html_block = f"<denk>Internet uplink succesvol.\nGezocht op: '{request.prompt}'\nGevonden info: {internet_info}</denk>\n"
        formatted_prompt = f"<|user|>\n[Context: {internet_info}]\n{request.prompt}\n<|assistant|>\n"
    else:
        formatted_prompt = f"<|user|>\n{request.prompt}\n<|assistant|>\n"

    # Tokenizen en sturen naar de GPU/CPU
    context = torch.tensor([tokenizer.encode(formatted_prompt).ids], dtype=torch.long, device=DEVICE)
    
    with torch.no_grad():
        generated_tokens = model.generate(
            context, 
            max_new_tokens=request.max_tokens, 
            tokenizer=tokenizer,
            temperature=request.temperature, 
            top_k=30
        )[0]
        full_output = tokenizer.decode(generated_tokens.tolist(), skip_special_tokens=False)
        
    # Sloop de prompt-tekst uit het uiteindelijke antwoord
    generated_text = full_output.replace(formatted_prompt, "").replace("<|end|>", "").strip()
    
    # --- DE ULTIEME ORBIT-SPACESTAR RETREATING & CLEANING FILTER ---
    # Sloop alle rare achtergrond-gedachten eruit en forceer zijn naam!
    
    # 1. Herken interne hersenspinsels en stop ze in jouw vette frontend <denk> blok!
    if any(trigger in generated_text for trigger in ["Discuss ramen", "User greets", "User asks", "Provide the basic", "Analyze the"]):
        parts = re.split(r'(Hello!|Hi!|I am|I think|To print|Based on)', generated_text, maxsplit=1)
        if len(parts) >= 3:
            internal_thought = parts[0].strip()
            actual_speech = parts[1] + parts[2]
            generated_text = f"<denk>Interne model-analyse:\n{internal_thought}</denk>\n{actual_speech}"
            
    # 2. KEIHARDE KORRECTIE: Hij is NIET Orbit Ultra, hij is Orbit-SpaceStar!
    generated_text = generated_text.replace("Orbit Ultra", "Orbit-SpaceStar")
    
    # 3. Voorkom dubbele loops op je scherm (als hij twee keer dezelfde vraag aan zichzelf stelt, knippen we hem af)
    if "What is the Python command" in generated_text:
        sub_parts = generated_text.split("What is the Python command")
        generated_text = sub_parts[0].strip()

    # Plak het internet-denkblok (als dat er is) aan het antwoord vast
    final_response = think_html_block + generated_text
    
    return {
        "version_used": request.version, 
        "response": final_response.strip()
    }