app.py

import gradio as gr
import os
from agent import analyze_case
from rag7 import generate_practical_guide
from rag24 import generate_academic_analysis
import dataset_manager

print("[DEBUG] Importing modules...")

HF_TOKEN = os.getenv("HF_TOKEN")
if not HF_TOKEN:
    raise RuntimeError("Missing HF_TOKEN in Secrets")

print("[DEBUG] HF_TOKEN found")

# Global variables for tracking
current_timestamp = None
current_input = None
current_acad_response = None
current_pract_response = None
current_metadata = None
current_temp_acad = None
current_topp_acad = None
current_context = None  # NEW: store citations to display

print("[DEBUG] Global variables initialized")

def generate_responses(
    message,
    temp_acad, topp_acad, max_tok_acad,
    temp_pract, topp_pract, max_tok_pract
):
    """
    Generates both responses (academic + practical) and enables rating system
    """
    global current_timestamp, current_input, current_acad_response, current_pract_response
    global current_metadata, current_temp_acad, current_topp_acad, current_context
    
    from datetime import datetime
    
    print(f"\n{'='*80}")
    print(f"📝 NEW QUERY: {message[:100]}...")
    print(f"{'='*80}")
    
    current_timestamp = datetime.now().isoformat()
    current_input = message
    current_temp_acad = temp_acad
    current_topp_acad = topp_acad
    
    try:
        # 1. ANALYZE CASE with RAG24 (retrieval + conflict detection)
        print("\n🔍 Step 1: Analyzing case with agent.analyze_case()...")
        result = analyze_case(message, k=5)
        metadata = result["metadata"]
        citations = result["citations"]
        
        current_metadata = metadata
        current_context = citations  # STORE for display
        
        print(f"✅ Analysis completed:")
        print(f"   - Conflict type: {metadata['conflict_type']} ({metadata['confidence']}%)")
        print(f"   - Sources retrieved: {metadata['num_sources']}")
        print(f"   - Unique tags: {metadata['unique_tags']}")
        
        # 2. GENERATE ACADEMIC ANALYSIS with Llama-3.3-70B-Instruct
        print("\n🧠 Step 2: Generating academic analysis with Llama-3.3-70B...")
        acad_response, acad_metadata = generate_academic_analysis(
            user_message=message,
            context=citations,  # citations is the formatted string
            metadata=metadata,
            temperature=temp_acad,
            top_p=topp_acad,
            max_tokens=max_tok_acad
        )
        current_acad_response = acad_response
        
        print(f"✅ Academic analysis generated ({len(acad_response)} chars)")
        print(f"   - Works cited: {acad_metadata.get('logical_works_used', [])}")
        
        # 3. GENERATE PRACTICAL GUIDE with GPT-OSS-20B
        print("\n📋 Step 3: Generating practical guide with GPT-OSS-20B...")
        pract = generate_practical_guide(
            message,
            temperature=temp_pract,
            top_p=topp_pract,
            max_tokens=max_tok_pract
        )
        current_pract_response = pract
        
        print(f"✅ Practical guide generated ({len(pract)} chars)")
        print(f"\n{'='*80}\n")
        
        # Enable rating system
        return (
            acad_response,
            pract,
            citations,  # NEW: show academic context
            gr.update(interactive=True),   # stars
            gr.update(interactive=True),   # coach_notes
            gr.update(interactive=True),   # submit_btn
            gr.update(value="")             # feedback_out
        )
        
    except Exception as e:
        error_msg = f"❌ Error in generate_responses: {str(e)}"
        print(error_msg)
        import traceback
        traceback.print_exc()
        
        return (
            f"Error generating academic analysis:\n{str(e)}",
            f"Error generating practical guide:\n{str(e)}",
            f"Error in retrieval:\n{str(e)}",  # context_out
            gr.update(interactive=False),
            gr.update(interactive=False),
            gr.update(interactive=False),
            gr.update(value=error_msg)
        )

def save_rating_to_dataset(stars, coach_notes):
    """
    Saves coach rating to HuggingFace Dataset for fine-tuning
    """
    global current_timestamp, current_input, current_acad_response, current_pract_response
    global current_metadata, current_temp_acad, current_topp_acad, current_context
    
    if not current_timestamp:
        return "⚠️ No active interaction to rate"
    
    if not stars:
        return "⚠️ Please select a rating (1-5 stars)"
    
    try:
        # Extract star number
        star_value = int(stars.split()[0])
        
        print(f"\n⭐ SAVING RATING: {star_value} stars")
        print(f"📝 Coach notes: {coach_notes[:100] if coach_notes else 'No notes'}...")
        
        # Extract categories used
        categories_used = list(current_metadata.get("faiss_weights_applied", {}).keys())
        
        # Create dataset entry
        data = dataset_manager.create_dataset_entry(
            timestamp=current_timestamp,
            user_input=current_input,
            stars=star_value,
            coach_notes=coach_notes or "",
            conflict_type=current_metadata.get("conflict_type", "unknown"),
            confidence=current_metadata.get("confidence", 0),
            num_sources=current_metadata.get("num_sources", 0),
            unique_tags=current_metadata.get("unique_tags", []),
            sources_raw=current_metadata.get("sources_raw", []),
            logical_works_used=current_metadata.get("logical_works_used", []),
            categories_used=categories_used,
            faiss_weights_applied=current_metadata.get("faiss_weights_applied", {}),
            temperature_used=current_temp_acad,
            top_p_used=current_topp_acad,
            academic_response=current_acad_response,
            practical_response=current_pract_response,
            citations_context=current_context or ""  # NEW: full academic context
        )
        
        # Push to HuggingFace
        dataset_manager.push_interaction_to_dataset(data)
        
        success_msg = f"✅ Rating saved successfully to HuggingFace ({star_value} stars)"
        print(success_msg)
        print(f"📊 Context saved: {len(current_context) if current_context else 0} chars")
        return success_msg
        
    except Exception as e:
        error_msg = f"❌ Error saving to dataset: {str(e)}"
        print(error_msg)
        import traceback
        traceback.print_exc()
        return error_msg

# ============================================================================
# GRADIO INTERFACE
# ============================================================================

print("[DEBUG] Creating Gradio interface...")

with gr.Blocks(title="Quoota RAG24 Alpha") as demo:
    
    gr.Markdown("""
    # 🧠 Quoota RAG24 Alpha
    ### Corporate Conflict Analysis System for Fine-Tuning
    
    **RAG24:** Academic analysis based on 24 works (16,997 vectorized chunks)  
    **RAG7:** Step-by-step practical guide
    """)
    
    with gr.Row():
        # LEFT COLUMN: Input and configuration
        with gr.Column(scale=2):
            inp = gr.Textbox(
                label="Describe the workplace conflict",
                placeholder="E.g.: My boss asks for updates every hour...",
                lines=4
            )
            btn = gr.Button("🔍 Analyze Conflict", variant="primary")
        
        # RIGHT COLUMN: LLM Parameters
        with gr.Column(scale=1):
            gr.Markdown("### ⚙️ RAG24 Parameters")
            gr.Markdown("*Llama-3.3-70B-Instruct (Academic)*")
            temp_acad = gr.Slider(0.1, 1.0, value=0.3, step=0.1, label="Temperature")
            topp_acad = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
            max_tok_acad = gr.Slider(256, 1024, value=768, step=128, label="Max tokens")
            
            gr.Markdown("### ⚙️ RAG7 Parameters")
            gr.Markdown("*GPT-OSS-20B (Practical)*")
            temp_pract = gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature")
            topp_pract = gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p")
            max_tok_pract = gr.Slider(256, 2048, value=2048, step=256, label="Max tokens")
            
            gr.Markdown("---")
            gr.Markdown("### ⭐ Rating System")
            stars = gr.Radio(
                choices=["1 star", "2 stars", "3 stars", "4 stars", "5 stars"],
                label="Response Quality",
                value=None,
                interactive=False
            )
            coach_notes = gr.Textbox(
                label="Coach Notes (optional)",
                placeholder="E.g.: Excellent EGO conflict detection, cited Brown & Levinson correctly...",
                lines=3,
                interactive=False
            )
            submit_btn = gr.Button(
                "💾 Save to HuggingFace",
                interactive=False,
                variant="secondary"
            )
            feedback_out = gr.Textbox(
                label="Status",
                interactive=False,
                show_label=False
            )
    
    # ROW 1: Main outputs (analysis + guide)
    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown("### 📚 Academic Analysis (RAG24)")
            gr.Markdown("*Based on 24 academic works with multilingual system*")
            acad_out = gr.Textbox(
                lines=25,
                max_lines=40,
                interactive=False
            )
        
        with gr.Column(scale=2):
            gr.Markdown("### 📋 Practical Guide (RAG7)")
            gr.Markdown("*Concrete steps to resolve the conflict*")
            pract_out = gr.Textbox(
                lines=25,
                max_lines=40,
                interactive=False
            )
    
    # ROW 2: Academic context (collapsible)
    with gr.Row():
        with gr.Column():
            with gr.Accordion("🔬 Academic Context (for coach validation)", open=True):
                gr.Markdown("""
                **This panel shows the context the LLM received to generate the academic analysis.**  
                
                Use it to validate if:
                - ✅ Retrieved sources are relevant to the conflict
                - ✅ Methodological weights are correct (ego/data/mixed)
                - ✅ Hybrid scores make sense (similarity + weight + tags)
                - ✅ The LLM cited sources correctly in its academic response
                
                **Included information:**
                - Applied retrieval system (conflict type, weights by category)
                - Top 5 sources with broken-down hybrid scoring
                - Full content of each chunk
                - Authors identified in each source
                """)
                context_out = gr.Textbox(
                    label="Full Academic Citations (sent to LLM)",
                    lines=30,
                    max_lines=50,
                    interactive=False
                )
    
    # EXAMPLES
    gr.Examples(
        examples=[
            ["My department colleague always takes credit for my ideas in meetings. She says 'we thought' but the ideas are mine. It's happened three times this month and I'm fed up."],
            ["My boss is never in the office and keeps calling to demand work without providing the means. He asks for reports that require access to databases only he can authorize. I'm exhausted."],
            ["In team meetings, my boss interrupts me every time I speak. He doesn't let me finish a sentence. Others don't even look at me when I try to intervene. I feel invisible."],
            ["A colleague made a mistake in a report I reviewed. Now everyone thinks the mistake was mine because I signed it. They avoid me in the hallways. No one talks to me."],
            ["My boss assigns me twice as many tasks as my colleague. When I ask why, he says 'because you're faster'. But that means I work twice as hard for the same salary. It's unfair."],
            ["My boss asks for updates every hour on what I'm doing. She messages to ask 'what are you working on' constantly. I can't focus. I feel like she doesn't trust me."],
            ["In yesterday's meeting, my boss said in front of everyone 'even an intern could do this'. He was talking about MY presentation. I was paralyzed. I didn't know what to say."],
            ["My boss moved me to a different project without telling me. I found out because a colleague asked why I wasn't in the meeting anymore. I'd been on that project for 6 months. I feel disregarded."],
            ["Someone has been saying I leaked confidential information. It's a lie. Now my boss looks at me strangely and I was excluded from the last important meeting. I don't know who said it or how to prove it."],
            ["I've been working until 11pm every day for 3 months. I asked my boss for help a month ago. He said 'hang in there a bit longer'. I can't take it anymore. I'm on the verge of collapse. I don't sleep. I don't eat well. I feel like I'm going to explode."]
        ],
        inputs=inp,
        label="💡 Example Cases"
    )
    
    # EVENTS
    btn.click(
        fn=generate_responses,
        inputs=[
            inp,
            temp_acad, topp_acad, max_tok_acad,
            temp_pract, topp_pract, max_tok_pract
        ],
        outputs=[
            acad_out,
            pract_out,
            context_out,  # NEW: show academic context
            stars,
            coach_notes,
            submit_btn,
            feedback_out
        ]
    )
    
    submit_btn.click(
        fn=save_rating_to_dataset,
        inputs=[stars, coach_notes],
        outputs=[feedback_out]
    )

print("[DEBUG] Gradio interface created successfully")

# ============================================================================
# LAUNCH
# ============================================================================

if __name__ == "__main__":
    print("\n" + "="*80)
    print("🚀 LAUNCHING QUOOTA RAG24 ALPHA")
    print("="*80 + "\n")
    demo.launch()
    print("\n✅ Demo launched successfully\n")