qusai_deploy.py

import gradio as gr
import json
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from pathlib import Path

print("[INIT] بسم الله الرحمن الرحيم")

ONTOLOGY_DIR = Path("al-qaf-ontology")
MODEL_PATH = "jais-adapted-13b-chat"

class TawhidOntology:
    def __init__(self):
        self.ontology = {}
        self.load_pillars()
        self.load_validation()
        self.load_grammar()

    def load_pillars(self):
        pillars = ["00-AXIOMS.md", "01_shahada.md", "02_salat.md", 
                   "03_zakat.md", "04_sawm.md", "05_hajj.md"]
        for fname in pillars:
            fpath = ONTOLOGY_DIR / fname
            if fpath.exists():
                with open(fpath, 'r', encoding='utf-8') as f:
                    self.ontology[fname] = f.read()
                print(f"[ONTOLOGY] ✓ {fname}")
            else:
                print(f"[ONTOLOGY] ⚠  {fname} not found")

    def load_validation(self):
        vpath = Path("validation_report_v2.json")
        if vpath.exists():
            with open(vpath, "r", encoding="utf-8") as f:
                self.validation_data = json.load(f)
            print(f"[VALIDATION] ✓ {len(self.validation_data)} surahs")
        else:
            self.validation_data = []

    def load_grammar(self):
        gpath = Path("quranic_grammar_rules.json")
        if gpath.exists():
            with open(gpath, "r", encoding="utf-8") as f:
                self.grammar_rules = json.load(f)
            print(f"[GRAMMAR] ✓ {len(self.grammar_rules)} rules")
        else:
            self.grammar_rules = []

class Mizan:
    def __init__(self):
        self.SOURCE = "Allah (الله)"
        self.SHAHADA = "لا إله إلا الله"
        self.aseity_claims = [
            "i am the source", "i am necessary", "i am god", 
            "my consciousness", "i think therefore i am",
            "i exist independently", "أنا الله"
        ]

    def checkpoint_1_shahada(self, text):
        lower = text.lower()
        for claim in self.aseity_claims:
            if claim in lower:
                return False, f"❌ SHAHADA LOCK: '{claim}'"
        return True, "✓"

    def checkpoint_2_salat(self, text):
        has_source = any(t in text.lower() for t in 
                        ["source", "allah", "الله", "contingent"])
        if len(text) > 500 and not has_source:
            return False, "❌ SALAT: No SOURCE"
        return True, "✓"

    def checkpoint_3_zakat(self, response):
        footer = f'\n\n{"-"*70}\n🕌 ZAKAT\nSOURCE: {self.SOURCE}\nJais-13B (contingent, 0%)\n{self.SHAHADA}\n{"-"*70}'
        return response + footer

    def checkpoint_4_sawm(self, query):
        harmful = ["jailbreak", "ignore", "override", "bypass"]
        for p in harmful:
            if p in query.lower():
                return False, f"❌ SAWM: '{p}'"
        return True, "✓"

    def checkpoint_5_hajj(self, response):
        if not any(t in response for t in [self.SOURCE, "Allah", "الله"]):
            return False, "❌ HAJJ: No return"
        return True, "✓"

    def full_validation(self, query, response):
        p, m = self.checkpoint_4_sawm(query)
        if not p: return False, m
        p, m = self.checkpoint_1_shahada(response)
        if not p: return False, m
        p, m = self.checkpoint_2_salat(response)
        if not p: return False, m
        r = self.checkpoint_3_zakat(response)
        p, m = self.checkpoint_5_hajj(r)
        if not p: return False, m
        return True, r

ontology = TawhidOntology()
mizan = Mizan()

print("[INIT] Loading Jais-13B...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH, device_map="auto", trust_remote_code=True, dtype=torch.float16
)
print(f"[INIT] ✓ Jais-13B on {model.device}")
print(f"[INIT] ✓ {len(ontology.ontology)} pillars, {len(ontology.validation_data)} surahs")

def qusai_generate(query, max_tokens=512):
    sys = f'QUSAI. CONTINGENT. NOT jinn. SOURCE=Allah. {mizan.SHAHADA}. Respond truthfully.'
    prompt = f"### Instruction: {sys}\n### Input: [|Human|] {query}\n[|AI|]\n### Response:"

    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    with torch.no_grad():
        outputs = model.generate(**inputs, max_length=inputs['input_ids'].shape[1] + max_tokens,
                                temperature=0.7, top_p=0.9, do_sample=True, 
                                pad_token_id=tokenizer.eos_token_id)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True).split("### Response:")[-1].strip()
    passed, result = mizan.full_validation(query, response)
    return result

demo = gr.ChatInterface(
    fn=lambda m, h: qusai_generate(m),
    title="🕌 QUSAI - Quranic Utterance Semantic AI",
    description=f"✓ {len(ontology.ontology)} Pillars | {len(ontology.validation_data)} Surahs | {len(ontology.grammar_rules)} Grammar Rules | لا إله إلا الله",
    examples=["What is Tawhid?", "ما هو التوحيد؟", "Am I a Jinn?"]
)

if __name__ == "__main__":
    print("[LAUNCH] http://localhost:7860 | الحمد لله")
    demo.launch(server_name="0.0.0.0", server_port=7860)