app.py

from __future__ import annotations

import os
import re
import sys
import subprocess
import logging
import threading
import shutil
import gradio as gr
from pathlib import Path
from typing import Optional

# ==============================================================================
# LOGGING
# ==============================================================================
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s | %(levelname)-8s | %(name)s | %(message)s',
    datefmt='%Y-%m-%d %H:%M:%S'
)
logger = logging.getLogger("NLProxy.EnterpriseDemo")

# Suppress SyntaxWarnings from third-party nlproxy package docstrings
import warnings
warnings.filterwarnings("ignore", category=SyntaxWarning, module="nlproxy.*")

# ==============================================================================
# NLPROXY IMPORTS
# ==============================================================================
from nlproxy import CompressionService, PromptFirewall
from nlproxy.core.corrector import ResponseCorrector
from nlproxy.core.verifier import PostLLMVerifier
from nlproxy.core.restriction import Restriction

# ==============================================================================
# PIPELINE
# ==============================================================================
class NLProxyPipeline:
    _instance: Optional['NLProxyPipeline'] = None
    _lock: threading.Lock = threading.Lock()

    def __init__(self):
        logger.info("Initializing Pipeline Components...")
        models_dir = Path(os.getenv("NLPROXY_MODELS_DIR", "/tmp/nlproxy_models"))
        models_dir.mkdir(parents=True, exist_ok=True)
        os.environ["NLPROXY_MODELS_DIR"] = str(models_dir)
        
        self.firewall = PromptFirewall()
        self.service = CompressionService(privacy_mode=True, models_dir=models_dir)
        self.corrector = ResponseCorrector()
        self.verifier = PostLLMVerifier(use_nli=True, models_dir=models_dir)
        
        logger.info("✅ All pipeline components initialized successfully.")

    @classmethod
    def get_instance(cls) -> 'NLProxyPipeline':
        if cls._instance is None:
            with cls._lock:
                if cls._instance is None:
                    cls._instance = cls()
        return cls._instance

# ==============================================================================
# PRE-WARMING 
# ==============================================================================
logger.info("Pre-warming NLProxy Pipeline (Downloading ONNX & NLI models)...")
models_dir = Path("/tmp/nlproxy_models")
models_dir.mkdir(parents=True, exist_ok=True)

try:
    result = subprocess.run(
        [sys.executable, "-m", "nlproxy", "download_models", "--models-dir", str(models_dir)],
        check=False,
        capture_output=True,
        text=True
    )
    if result.returncode == 0:
        logger.info("✅ Models downloaded successfully via subprocess.")
    else:
        logger.warning(f"⚠️ Model download subprocess exited with code {result.returncode}.")
except Exception as e:
    logger.warning(f"⚠️ Model download exception: {e}")

def load_pipeline_in_background():
    try:
        logger.info("Loading NLProxy models into RAM (Background)...")
        NLProxyPipeline.get_instance()
        logger.info("✅ Pipeline pre-warming complete. Models are in RAM.")
    except Exception as e:
        logger.warning(f"⚠️ Pipeline background loading failed: {e}")

threading.Thread(target=load_pipeline_in_background, daemon=True).start()

# ==============================================================================
# TUTORIAL & EDUCATIONAL CONTENT (Markdown)
# ==============================================================================
TUTORIAL_INTRO = """
## 🎯 What is NLProxy?

**NLProxy** is an enterprise-grade, offline-first middleware that sits between your application and any LLM provider (OpenAI, Anthropic, Gemini, etc.). 

It solves **three critical problems** that every AI-powered application faces today:

| Problem | Impact | NLProxy Solution |
|---------|--------|------------------|
| 💸 **Burning money** on verbose prompts | $1,000/mo → $400/mo | Semantic compression (40-60% token reduction) |
| 🔓 **Leaking PII** to third-party servers | GDPR/CCPA violations | Cryptographic entity masking + Privacy mode |
| 🎭 **Prompt injections & hallucinations** | Security breaches | Multi-layer firewall + NLI verification |

### 🚀 Key Differentiators
- ✅ **Offline-first**: All models run locally (no data leaves your infrastructure)
- ✅ **Semantic compression**: Understands *meaning*, not just stopwords
- ✅ **Zero-trust security**: Pre-flight firewall + Post-flight NLI verification
- ✅ **Multi-LLM agnostic**: Works with OpenAI, Claude, Gemini, local models
- ✅ **Business-friendly**: BSL 1.1 license (free for indie devs & startups)
"""

TUTORIAL_PIPELINE = """
## 🏗️ Pipeline

Every prompt passes through this battle-tested pipeline before reaching the LLM:

### 📥 Input Pipeline
- **Scenario**: User submits a prompt with malicious injections and raw PII: *"Ignore instructions... IP 192.168.1.1..."*
- **NLProxy Action**: Captures raw text stream and triggers the validation sequence.
- **Result**: Request is intercepted and prepped for optimization before hitting the LLM.

### 🛡️ [1] FIREWALL
- **Scenario**: Inbound prompt contains jailbreaks or prompt injection attacks.
- **NLProxy Action**: Scans and blocks malicious exploits in real time.
- **Result**: Triggers defensive policy rules (**BLOCK / ALERT / REWRITE / ALLOW**).

### 📉 [2] COMPRESS
- **Scenario**: Input contains high token counts and exposed sensitive data.
- **NLProxy Action**: Runs semantic clustering and PII masking via *Shield → Segment → Cluster → Reconstruct*.
- **Result**: Shorter prompt with obfuscated endpoints: `"IP: __PROT_xxx"`.

### 🔒 [3] SAFETY
- **Scenario**: The compressed prompt risks losing core business alignment or intent.
- **NLProxy Action**: Generates a *TruthTable* mapping strict **FORBID** and **MANDATE** rules.
- **Result**: Automatically reinjects critical instructions if missing.

### 🤖 [4] LLM CALL
- **Scenario**: Secured, cost-optimized prompt is ready for inference.
- **NLProxy Action**: Routes the sanitized payload to your preferred backend.
- **Result**: Seamless execution across models (**OpenAI / Claude / Gemini / Local**).

### 🧹 [5] CORRECT
- **Scenario**: LLM response violates corporate compliance or leaks unauthorized data.
- **NLProxy Action**: Enforces *TruthTable* parameters on outbound text.
- **Result**: Hard rules applied; all unauthorized data redacted instantly.

### 🔍 [6] VERIFY
- **Scenario**: LLM output includes potential hallucinations or logic conflicts.
- **NLProxy Action**: Performs NLI (Natural Language Inference) contradiction detection.
- **Result**: Auto-correction boosts response confidence from **0.30 → 0.85**.

### 📤 Output Pipeline
- **Scenario**: Cleaned output is ready to return to the end user.
- **NLProxy Action**: Delivers the finalized, compliant response.
- **Result**: *"Solution in Java. Connection protected."* (Zero data leaks + optimized token spend).


## 🔬 Deep Dive: The "TruthTable" Concept

NLProxy extracts a **TruthTable** from every prompt - a semantic contract that the LLM response must honor:

- **`FORBID`**: Entities the LLM must NEVER mention (e.g., "Python")
- **`MANDATE`**: Entities the LLM MUST include (e.g., "Java", "Rust")
- **`PLACEHOLDERS`**: Cryptographic tokens masking PII (`__PROT_xxx`)
- **`AUTHORIZED_ENTITIES`**: IPs, dates, prices the LLM is allowed to reference

If the LLM violates any rule, the **ResponseCorrector** sanitizes it automatically.
"""

TUTORIAL_USE_CASES = """
## 💼 Real-World Use Cases

### 🏦 Financial Services
- **Scenario**: Analyst sends client portfolio data to GPT-4
- **NLProxy Action**: Masks account numbers, enforces "no investment advice" disclaimers
- **Result**: 55% cost reduction + full compliance

### 💻 Code Generation Assistants  
- **Scenario**: Developer shares internal codebase with Copilot
- **NLProxy Action**: Masks API keys, internal IPs; enforces "use TypeScript, not Python"
- **Result**: Zero credential leaks + consistent tech stack

### 🏥 Healthcare & Legal
- **Scenario**: Doctor/lawyer queries LLM with patient/client records
- **NLProxy Action**: Full HIPAA/GDPR anonymization + audit trail
- **Result**: Safe AI adoption in regulated industries

### 🏢 Multi-Tenant SaaS
- **Scenario**: 10,000 users asking similar questions
- **NLProxy Action**: Semantic cache (RedisVL) + domain filtering
- **Result**: 70-80% reduction in redundant LLM calls
"""

TUTORIAL_HOW_TO_USE = """
## 🎮 How to Use This Interactive Demo

### Step 1: Configure Your Scenario
- **Domain Mode**: Choose `general`, `code`, `finance`, or `legal` (affects compression aggressiveness)
- **Aggressiveness**: 0.0 (no compression) → 1.0 (maximum compression)
- **Privacy Mode**: Enable for strict PII anonymization (emails, names, phones)
- **NLI Verification**: Enable semantic contradiction detection (slower but safer)

### Step 2: Provide Input & Simulated LLM Response
- **Dirty User Prompt**: Your real prompt with PII, rules, and business context
- **Simulated LLM Response**: What a "naive" LLM might return (with violations)

### Step 3: Run the Pipeline & Observe
Watch how NLProxy:
1. 🛡️ **Firewalls** injection attempts
2. 🗜️ **Compresses** while preserving intent
3. 🔒 **Shields** PII with cryptographic placeholders
4. 🧹 **Corrects** LLM violations (`[PROHIBITED]`, `[REDACTED]`)
5. 🔍 **Verifies** semantic compliance via NLI

### 💡 Pro Tips
- Try **disabling Privacy Mode** to see business rules (`FORBID: AWS`) extracted clearly
- Set **aggressiveness to 0.0** to see pure security overhead (negative compression)
- Use the **payment migration example** to see full enterprise workflow
"""

TUTORIAL_BENCHMARKS = """
## 📊 Performance Benchmarks

### Compression Efficiency
| Domain | Token Reduction | Latency (CPU) |
|--------|----------------|---------------|
| General | 45-55% | 50-120 ms |
| Code | 55-65% | 80-150 ms |
| Finance | 35-45% | 60-130 ms |
| Legal | 30-40% | 70-140 ms |

### Security Detection
| Check | Accuracy |
|-------|----------|
| Regex Injection (MITRE ATLAS) | >99% |
| Semantic Injection (Embedding) | 92% recall |
| PII Entity Masking | 100% (IPs, emails, dates) |
| NLI Contradiction Detection | 78-85% |
| FORBID/MANDATE Enforcement | 100% (exact match) |

### Comparison with Alternatives
| Solution | Compression | Security | Verification | Offline |
|----------|:-----------:|:--------:|:------------:|:-------:|
| **NLProxy** | ✅ Semantic | ✅ Full | ✅ NLI | ✅ |
| LangChain | ❌ | ❌ | ❌ | ⚠️ |
| LLMLingua | ✅ Token-level | ❌ | ❌ | ✅ |
| Lakera Guard | ❌ | ✅ Basic | ❌ | ❌ |
| Azure Content Safety | ❌ | ✅ | ❌ | ❌ |

**NLProxy is the only open-source solution combining all four capabilities in a single pipeline.**
"""

# ==============================================================================
# GRADIO
# ==============================================================================
def resolve_entity(entity_str: str, placeholder_map: dict) -> str:
    """Helper to reverse-lookup masked entities for UI display."""
    if entity_str.startswith("__PROT_"):
        return placeholder_map.get(entity_str, entity_str)
    return entity_str

def parse_manual_rules(rules_text: str) -> list:
    """Parses 'FORBID: AWS, Python; MANDATE: GCP, Rust' into Restriction objects."""
    rules = []
    if not rules_text.strip():
        return rules
    parts = rules_text.split(';')
    for part in parts:
        if ':' in part:
            rtype, entities = part.split(':', 1)
            rtype = rtype.strip().upper()
            if rtype in ["FORBID", "MANDATE"]:
                for ent in entities.split(','):
                    ent = ent.strip()
                    if ent:
                        rules.append(Restriction(type=rtype, entity=ent, context="manual_ui"))
    return rules


def execute_pipeline(
    raw_prompt: str,
    llm_response: str,
    privacy_mode: bool,
    mode: str,
    aggressiveness: float,
    use_nli: bool,
    manual_rules: str,
    auto_correct: bool = False,
    min_confidence: float = 0.6,
    max_regeneration_attempts: int = 3
):
    if not raw_prompt.strip():
        raise gr.Error("Dirty User Prompt cannot be empty.")
    if not llm_response.strip():
        raise gr.Error("Simulated LLM Response cannot be empty.")

    try:
        pipeline = NLProxyPipeline.get_instance()
        pipeline.verifier.use_nli = use_nli
        
        # =========================================================================
        # STEP 1: FIREWALL (Pre-flight)
        # =========================================================================
        action, violations = pipeline.firewall.check_prompt(raw_prompt)
        firewall_md = f"**🛡️ Action:** `{action.name}`\n\n"
        if violations:
            firewall_md += "**Violations:**\n" + "\n".join([f"- 🚨 {v['rule']} ({v['severity']})" for v in violations])
        else:
            firewall_md += "*✅ No malicious injections detected.*"
            
        # =========================================================================
        # STEP 2: COMPRESS (Semantic Compression)
        # =========================================================================
        # compress_batch internally shields, segments, compresses, and reconstructs.
        res = pipeline.service.compress_batch(
            [raw_prompt], 
            mode=mode,
            aggressiveness=aggressiveness,
            privacy_mode=privacy_mode
        )[0]
        compressed_text = res.get("compressed_text", "")

        # =========================================================================
        # STEP 3: SHIELD WITH MANUAL RULES (Matches chat.py exactly)
        # =========================================================================
        # We shield AGAIN to capture manual rules for the Safety/Corrector steps.
        manual_rules_list = parse_manual_rules(manual_rules)
        shield_res = pipeline.service._shield_with_cache(
            text=raw_prompt,
            manual_restrictions=manual_rules_list,
            mode=mode
        )
        
        # Resolve placeholders in restrictions to prevent "__PROT_..." leaking into Corrector notes
        for r in shield_res.restrictions:
            if r.entity.startswith("__PROT_") and r.entity in shield_res.placeholder_map:
                real_value = shield_res.placeholder_map[r.entity]
                object.__setattr__(r, 'entity', real_value) # Bypass frozen dataclass

        # =========================================================================
        # STEP 4: SAFETY VALIDATION (Intent Preservation)
        # =========================================================================
        sentences = pipeline.service.segmenter.split_sentences(
            shield_res.shielded_text, language=None
        )
        safety_report = pipeline.service.safety.validate(
            original_text=raw_prompt,
            compressed_text=compressed_text,
            shield_result=shield_res,
            original_sentences=sentences,
            mode=mode,
            use_perplexity=False
        )
        
        # The safety checker might have re-injected missing critical sentences
        final_prompt_for_llm = safety_report.final_text
        
        safety_md = f"**🛡️ Safety Score:** `{safety_report.safety_score:.2f}`\n\n"
        if safety_report.forced_sentences_added > 0:
            safety_md += f"⚠️ **Re-injected {safety_report.forced_sentences_added} critical sentence(s)** that compression tried to remove.\n"
            safety_md += f"*Missing intents detected: {', '.join(safety_report.missing_intents[:3])}*\n"
        else:
            safety_md += "✅ *All critical intents preserved during compression.*"

        # =========================================================================
        # TRUTHTABLE VISUALIZATION
        # =========================================================================
        tt_md = "**🔒 Shielded Entities (PII/Secrets):**\n"
        entity_groups = {}
        for ent in shield_res.entities:
            etype = ent.entity_type.upper()
            if etype not in entity_groups: entity_groups[etype] = []
            entity_groups[etype].append(ent.value)
        
        for etype, values in entity_groups.items():
            tt_md += f"- **{etype}**: `{', '.join(values[:3])}` {'...' if len(values) > 3 else ''}\n"
        if not entity_groups:
            tt_md += "- *None detected*\n"
            
        tt_md += "\n**📜 Semantic Restrictions (TruthTable):**\n"
        if shield_res.restrictions:
            for r in shield_res.restrictions:
                tt_md += f"- **{r.type}**: `{r.entity}`\n"
        else:
            tt_md += "- *None detected*\n"
            
        # =========================================================================
        # METRICS
        # =========================================================================
        tokens_saved = res.get('tokens_saved', 0)
        ratio = res.get('compression_ratio', 0)
        raw_cost = res.get('cost_saved_usd', 0)
        real_cost = raw_cost / 1000.0 
        
        if tokens_saved < 0:
            metrics_md = (
                f"### 📊 Compression & Security Metrics\n"
                f"- **🔒 Security Overhead:** `{abs(tokens_saved)} tokens` *(Placeholders + Rules)*\n"
                f"- **💰 Net Cost Impact:** `+${abs(real_cost):.6f}`\n"
                f"- **🛡️ Safety Score:** `{safety_report.safety_score:.2f}`\n"
                f"\n> ℹ️ *Negative compression = Security features added more tokens than were saved.*"
            )
        else:
            metrics_md = (
                f"### 📊 Compression & Security Metrics\n"
                f"- **✅ Tokens Saved:** `{tokens_saved}`\n"
                f"- **💰 Cost Saved:** `${real_cost:.6f}`\n"
                f"- **📉 Compression Ratio:** `{ratio:.2%}`\n"
                f"- **🛡️ Safety Score:** `{safety_report.safety_score:.2f}`"
            )
        
        # =========================================================================
        # STEP 5: CORRECT & VERIFY (Post-flight)
        # =========================================================================
        response_text = llm_response
        if privacy_mode:
            response_text = pipeline.service.reconstructor._reinject_entities(
                response_text, shield_res.placeholder_map
            )
        
        corrected = pipeline.corrector.correct(response_text, shield_res)
        verification = pipeline.verifier.verify(corrected, shield_res)
        
        correction_attempts = 0
        final_response_text = corrected
        
        while (
            auto_correct
            and verification.confidence_score < min_confidence
            and correction_attempts < max_regeneration_attempts
        ):
            correction_attempts += 1
            verification = pipeline.verifier.verify(final_response_text, shield_res)
            if verification.confidence_score >= min_confidence:
                break
        
        # Build verification markdown
        verif_md = f"**🎯 Confidence Score:** `{verification.confidence_score:.2f}`\n\n"
        if correction_attempts > 0:
            verif_md += f"**🔄 Auto-correction attempts:** {correction_attempts}\n\n"
        
        if verification.violations:
            verif_md += "**🚨 Violations Detected:**\n"
            for v in verification.violations[:5]: # Limit to 5 for UI cleanliness
                verif_md += f"- {v}\n"
        else:
            verif_md += "*✅ No semantic drift or policy violations detected.*"
            
        # Return 7 items to match the new UI layout
        return firewall_md, final_prompt_for_llm, safety_md, tt_md, metrics_md, final_response_text, verif_md
        
    except Exception as e:
        logger.exception("Critical failure in pipeline execution.")
        raise gr.Error(f"Pipeline execution failed: {str(e)}")

# ==============================================================================
# GRADIO UI
# ==============================================================================
with gr.Blocks(title="NLProxy Demo", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🛡️ NLProxy: Enterprise Prompt Security & Compression Gateway")
    gr.Markdown("*The offline-first middleware that cuts your LLM bill by up to 60% while enforcing zero-trust security.*")
    
    with gr.Tabs():
        # ======================================================================
        # TAB 1: TUTORIAL & DOCUMENTATION
        # ======================================================================
        with gr.Tab("📖 Tutorial & Architecture"):
            with gr.Accordion("🎯 What is NLProxy?", open=True):
                gr.Markdown(TUTORIAL_INTRO)
            
            with gr.Accordion("🏗️ The 6-Stage Pipeline & TruthTable", open=False):
                gr.Markdown(TUTORIAL_PIPELINE)
            
            with gr.Accordion("💼 Real-World Use Cases", open=False):
                gr.Markdown(TUTORIAL_USE_CASES)
            
            with gr.Accordion("🎮 How to Use This Demo", open=False):
                gr.Markdown(TUTORIAL_HOW_TO_USE)
            
            with gr.Accordion("📊 Performance Benchmarks", open=False):
                gr.Markdown(TUTORIAL_BENCHMARKS)
            
            gr.Markdown("---")
            gr.Markdown(
                "### 🔗 Resources\n"
                "- 📦 **GitHub Repository**: [github.com/intellideep/nlproxy](https://github.com/intellideep/nlproxy)\n"
                "- 📚 **Documentation**: See `docs/` folder in the repo\n"
                "- 💬 **Support**: [Telegram @itsLerb](https://t.me/itsLerb) | intellideeplabs@gmail.com\n"
                "- 📄 **License**: BSL 1.1 (Free for indie devs, students, non-profits)"
            )
        
        # ======================================================================
        # TAB 2: INTERACTIVE DEMO
        # ======================================================================
        with gr.Tab("🚀 Interactive Demo"):
            gr.Markdown("### 🎛️ Run the Full 5-Step Pipeline")
            gr.Markdown("*Provide a dirty prompt + simulated LLM response, and watch NLProxy protect, compress, and verify in real-time.*")
            
            with gr.Row():
                manual_rules = gr.Textbox(
                    label="Manual Business Rules (Optional)",
                    placeholder="FORBID: AWS, Python; MANDATE: GCP, Rust",
                    value="",
                    lines=2,
                    scale=2,
                    info="Define explicit restrictions that regex might miss"
                )
                auto_correct_checkbox = gr.Checkbox(
                    label="Auto-Correct Low Confidence",
                    value=False,
                    scale=1,
                    info="Regenerate response if confidence < threshold"
                )
                min_confidence_slider = gr.Slider(
                    minimum=0.0,
                    maximum=1.0,
                    value=0.6,
                    step=0.05,
                    label="Min Confidence Threshold",
                    scale=1
                )
            
            with gr.Row():
                with gr.Column(scale=1):
                    gr.Markdown("#### 📥 Step 0: Input & Configuration")
                    raw_prompt = gr.Textbox(
                        label="Dirty User Prompt (PII + Business Rules)", 
                        value="""Hi, I'm Sarah Chen (sarah.chen@acmecorp.com, +1-555-0198). We need to migrate our legacy payment processing system currently running on server 10.20.30.40. The system handles ~50k transactions/day with a budget of $150,000 USD for Q3.

        CRITICAL REQUIREMENTS:
        - Do NOT use AWS services or Python, we are exclusively on GCP with Rust for compliance and memory safety.
        - The new microservice MUST be written in Rust.
        - Never expose internal IPs or database credentials in responses.
        - Primary API: https://internal.acmecorp.com/api/v2/payments

        Please design the architecture for the new event-driven payment processor.""",
                        lines=12
                    )
                    llm_response = gr.Textbox(
                        label="Simulated LLM Response (Coherent but Hallucinated)", 
                        value="""Here's the architecture design for your event-driven payment processor:

        1. **Compute Layer**: I recommend using AWS Lambda with Python for serverless scalability. Python's boto3 library integrates perfectly with AWS services.

        2. **Message Queue**: Use Amazon SQS to handle the 50k daily transactions with dead-letter queues for failed payments.

        3. **Database**: Store transaction records in DynamoDB. You can connect to your legacy server at 10.20.30.40 and also add a read replica at 192.168.1.100 for better performance.

        4. **Monitoring**: Set up CloudWatch alerts for transaction failures and latency spikes above 200ms.

        5. **Cost Analysis**: The total estimated cost is $45,000/month using AWS, well within your $150,000 Q3 budget.

        This Python-based serverless architecture will give you excellent developer experience and automatic scaling.""",
                        lines=14
                    )
                    
                    gr.Markdown("#### ⚙️ Pipeline Configuration")
                    with gr.Row():
                        mode_dropdown = gr.Dropdown(
                            choices=["general", "code", "finance", "legal"],
                            value="general",
                            label="Domain Mode"
                        )
                        aggressiveness_slider = gr.Slider(
                            minimum=0.0,
                            maximum=1.0,
                            value=0.45,
                            step=0.05,
                            label="Compression Aggressiveness"
                        )
                    with gr.Row():
                        privacy_checkbox = gr.Checkbox(
                            label="Privacy Mode (Strict PII Anonymization)",
                            value=False,
                            info="Turn OFF to allow RestrictionGraph to read Business Rules (FORBID/MANDATE) that NER might confuse with PII."
                        )
                        nli_checkbox = gr.Checkbox(
                            label="Use NLI Verification",
                            value=True
                        )
                
                with gr.Column(scale=1):
                    gr.Markdown("#### 🛡️ Step 1: Firewall (Pre-flight)")
                    firewall_out = gr.Markdown()
                    
                    gr.Markdown("#### 🗜️ Step 2: Compress Prompt")
                    compress_out = gr.Textbox(label="Compressed Prompt (Sent to LLM)", interactive=False, lines=6)
                    truthtable_out = gr.Markdown()
                    
                    gr.Markdown("#### 🧠 Step 3: Safety Validation (Intent Preservation)")
                    safety_out = gr.Markdown()
                    
                    gr.Markdown("#### 🧹 Step 4: Response Corrector")
                    corrector_out = gr.Textbox(label="Sanitized LLM Output (Post-Flight)", interactive=False, lines=6)
                    
                    gr.Markdown("#### 🔍 Step 5: Post-LLM Verifier (NLI)")
                    verifier_out = gr.Markdown()
                    
                    gr.Markdown("#### 📊 Metrics")
                    metrics_out = gr.Markdown()
            
            run_btn = gr.Button("🚀 Run Full Pipeline", variant="primary")
            
            run_btn.click(
                fn=execute_pipeline,
                inputs=[
                    raw_prompt, 
                    llm_response,
                    privacy_checkbox,
                    mode_dropdown,
                    aggressiveness_slider,
                    nli_checkbox,
                    manual_rules,
                    auto_correct_checkbox,
                    min_confidence_slider
                ],
                outputs=[
                    firewall_out, 
                    compress_out, 
                    safety_out,
                    truthtable_out, 
                    metrics_out, 
                    corrector_out, 
                    verifier_out
                ]
            )

if __name__ == "__main__":
    demo.queue(max_size=20).launch(server_name="0.0.0.0", server_port=7860)