app.py

import os
from typing import Any, Dict, List, Optional, Tuple

import gradio as gr
import httpx
import pandas as pd


# ─────────────────────────────────────────────────────────────────────────────
# Sample Data (Fallback when backend is unavailable)
# ─────────────────────────────────────────────────────────────────────────────

SAMPLE_LEDGER: List[Dict[str, Any]] = [
    {"concept": "Recursion base case", "status": "on_loan", "confidence": 0.92, "source_text": "Borrowed from AI explanation"},
    {"concept": "Binary search invariants", "status": "clear", "confidence": 0.88, "source_text": "Explained in own words"},
    {"concept": "Gradient descent", "status": "persists", "confidence": 0.79, "source_text": "Needs another clearing session"},
    {"concept": "RAG retrieval", "status": "on_loan", "confidence": 0.84, "source_text": "Added during prompt work"},
]

SAMPLE_HEATMAP = [
    {"concept": "Recursion base case", "count": 9},
    {"concept": "Binary search invariants", "count": 4},
    {"concept": "Gradient descent", "count": 7},
]

SAMPLE_METRICS = {
    "active": 2,
    "cleared": 1,
    "persists": 1,
    "debt_score": 67,
    "pending_sync": 0,
    "spoof_attempts": 3,
}


# ─────────────────────────────────────────────────────────────────────────────
# Backend Integration
# ─────────────────────────────────────────────────────────────────────────────

def backend_snapshot(base_url: str) -> Optional[Dict[str, Any]]:
    """Fetch live data from the knowledge backend API."""
    base = base_url.rstrip("/")
    try:
        with httpx.Client(timeout=6.0) as client:
            state = client.get(f"{base}/api/state", timeout=6.0).json()
            sync = client.get(f"{base}/api/sync/status", timeout=6.0).json()
            integrity = client.get(f"{base}/api/integrity/report", timeout=6.0).json()
            return {"state": state, "sync": sync, "integrity": integrity}
    except Exception:
        return None


def scout_demo(pasted_text: str) -> Tuple[str, str]:
    """Simulate Scout tagging concepts from pasted text."""
    if not pasted_text.strip():
        return "", "Paste some text to see Scout extract concepts."
    
    extracted_concepts = []
    keywords = ["recursion", "binary search", "gradient descent", "rag", "algorithm", "model", "training", "inference"]
    for keyword in keywords:
        if keyword.lower() in pasted_text.lower():
            extracted_concepts.append(keyword.capitalize())
    
    if not extracted_concepts:
        extracted_concepts = ["Learning (inferred from context)"]
    
    result = f"**Scout detected {len(extracted_concepts)} concept(s):**\n\n"
    for concept in extracted_concepts:
        result += f"- {concept} (confidence: ~85%)\n"
    
    status = f"✅ Tagged {len(extracted_concepts)} concept(s) → added to your ledger"
    return result, status


def sage_demo(concept: str, user_response: str) -> str:
    """Simulate Sage Socratic dialogue."""
    if not concept.strip():
        return "Enter a concept to start a clearing session."
    
    if not user_response.strip():
        return f"🦉 **Sage**: Let's talk about {concept}. Can you explain it in your own words?"
    
    quality_score = len(user_response.split()) / 10
    if quality_score < 2:
        return f"🦉 **Sage**: That's a start. But can you go deeper? What makes {concept} special or different?"
    elif quality_score < 4:
        return f"🦉 **Sage**: Good effort! Now, why would someone use {concept} in practice? When does it matter?"
    else:
        return f"✅ **Sage**: Excellent! You've clearly understood {concept}. This concept is now **CLEARED** in your ledger."


def get_ledger_display() -> pd.DataFrame:
    """Return the current ledger as a DataFrame for display."""
    backend_url = os.environ.get("KNOWLEDGE_API_URL", "").strip()
    snapshot = backend_snapshot(backend_url) if backend_url else None
    ledger = snapshot["state"]["debts"] if snapshot else SAMPLE_LEDGER
    
    df = pd.DataFrame(ledger)
    if not df.empty:
        df = df[["concept", "status", "confidence"]]
    return df


def get_metrics() -> Tuple[int, int, int, int]:
    """Return key metrics: active, cleared, persists, debt_score."""
    backend_url = os.environ.get("KNOWLEDGE_API_URL", "").strip()
    snapshot = backend_snapshot(backend_url) if backend_url else None
    ledger = snapshot["state"]["debts"] if snapshot else SAMPLE_LEDGER
    
    active = sum(1 for row in ledger if row.get("status") in {"on_loan", "persists"})
    cleared = sum(1 for row in ledger if row.get("status") in {"clear", "owned"})
    persists = sum(1 for row in ledger if row.get("status") == "persists")
    debt_score = round(((active + persists) / max(len(ledger), 1)) * 100) if ledger else 0
    
    return active, cleared, persists, debt_score


# ─────────────────────────────────────────────────────────────────────────────
# Gradio Interface
# ─────────────────────────────────────────────────────────────────────────────

with gr.Blocks(
    title="KnowLedge | Gemma 4 for Good",
    theme=gr.themes.Soft(
        primary_hue="slate",
        secondary_hue="amber",
    ),
) as demo:
    
    gr.Markdown(
        """
        # 🦉 KnowLedge
        
        **A local-first learning verification system for the Gemma 4 for Good hackathon.**
        
        KnowLedge turns pasted AI-assisted work into a guided mastery loop:
        - **Scout** extracts concepts from text
        - **Sage** clears them through Socratic dialogue
        - **Lens** verifies understanding with integrity checks
        - **Reports** share only anonymous aggregates with instructors
        """
    )
    
    with gr.Row():
        active, cleared, persists, debt_score = get_metrics()
        
        with gr.Column(scale=1):
            gr.Markdown(f"### 📊 Active\n\n**{active}** concepts on loan or persisting")
        with gr.Column(scale=1):
            gr.Markdown(f"### ✅ Cleared\n\n**{cleared}** owned concepts")
        with gr.Column(scale=1):
            gr.Markdown(f"### 🎯 Debt Score\n\n**{debt_score}%** (lower is better)")
        with gr.Column(scale=1):
            gr.Markdown(f"### 🛡️ Integrity\n\n**{SAMPLE_METRICS['spoof_attempts']}** spoof signals detected")
    
    with gr.Tabs():
        
        # ─────────────────────────────────────────────────────────────────────
        # Tab 1: Scout Demo
        # ─────────────────────────────────────────────────────────────────────
        with gr.TabItem("🔍 Scout — Extract Concepts"):
            gr.Markdown(
                """
                **Scout** automatically finds concepts inside pasted text. Try pasting an explanation or code snippet below.
                """
            )
            
            with gr.Row():
                with gr.Column(scale=2):
                    pasted_text = gr.Textbox(
                        label="Paste AI-assisted work or notes here",
                        placeholder="e.g., 'Recursion works by dividing a problem into smaller subproblems until reaching a base case...'",
                        lines=6,
                    )
                with gr.Column(scale=1):
                    scout_btn = gr.Button("🚀 Run Scout", size="lg")
            
            scout_output = gr.Markdown("Paste something to get started.")
            scout_status = gr.Textbox(label="Status", interactive=False, value="Ready.")
            
            scout_btn.click(
                scout_demo,
                inputs=[pasted_text],
                outputs=[scout_output, scout_status],
            )
        
        # ─────────────────────────────────────────────────────────────────────
        # Tab 2: Sage Demo
        # ─────────────────────────────────────────────────────────────────────
        with gr.TabItem("🦉 Sage — Socratic Clearing"):
            gr.Markdown(
                """
                **Sage** guides you through a Socratic clearing session. Pick a concept and explain it in your own words.
                """
            )
            
            concept_input = gr.Textbox(
                label="Concept to clear",
                placeholder="e.g., 'Binary Search'",
                value="Recursion",
            )
            
            response_input = gr.Textbox(
                label="Your explanation",
                placeholder="Explain the concept in your own words. Be as detailed as you can.",
                lines=4,
            )
            
            sage_btn = gr.Button("💭 Get Sage Response", size="lg")
            sage_output = gr.Markdown()
            
            sage_btn.click(
                sage_demo,
                inputs=[concept_input, response_input],
                outputs=[sage_output],
            )
        
        # ─────────────────────────────────────────────────────────────────────
        # Tab 3: Live Ledger
        # ─────────────────────────────────────────────────────────────────────
        with gr.TabItem("📖 Live Ledger"):
            gr.Markdown(
                """
                This is your concept ledger. Every concept you study is tracked here with its status:
                - **on_loan**: You pasted it but haven't cleared it yet.
                - **clear**: You explained it to Sage and passed.
                - **persists**: Lens found gaps, needs another session.
                """
            )
            
            ledger_df = get_ledger_display()
            ledger_table = gr.Dataframe(
                value=ledger_df,
                interactive=False,
                wrap=True,
            )
            
            refresh_btn = gr.Button("🔄 Refresh Ledger")
            refresh_btn.click(
                lambda: get_ledger_display(),
                outputs=[ledger_table],
            )
        
        # ─────────────────────────────────────────────────────────────────────
        # Tab 4: Architecture & Setup
        # ─────────────────────────────────────────────────────────────────────
        with gr.TabItem("🏗️ Architecture"):
            gr.Markdown(
                """
                ## System Design
                
                **Backend Stack:**
                - FastAPI for the web server
                - SQLite for concept ledger persistence
                - ChromaDB for curriculum context (RAG)
                - Ollama for local Gemma inference
                - Privacy-preserving sync with concept-level aggregates only
                
                **Key Features:**
                - Offline-first (no cloud dependency)
                - Session fingerprinting to detect gaming behavior
                - Anti-spoof scoring on Lens uploads
                - No student identifiers in instructor reports
                
                ## Quick Local Setup
                
                ```bash
                python3 -m venv .venv
                source .venv/bin/activate
                pip install -r requirements.txt
                
                # Terminal 1: Start Ollama
                ollama serve
                
                # Terminal 2: Start the backend
                python -m uvicorn knowledge.main:app --host 127.0.0.1 --port 8000
                
                # Optional: Load curriculum material
                python -m knowledge.vectorize path/to/course.pdf
                ```
                
                ## Deploy on Hugging Face Spaces
                
                1. Create a new Space with **Gradio** SDK
                2. Push this repo to the Space
                3. Set `KNOWLEDGE_API_URL` environment variable (optional, for live backend)
                4. Share the Space URL as your Kaggle demo link
                """
            )
        
        # ─────────────────────────────────────────────────────────────────────
        # Tab 5: For Judges
        # ─────────────────────────────────────────────────────────────────────
        with gr.TabItem("🎯 For Kaggle Judges"):
            gr.Markdown(
                """
                ## What Makes KnowLedge Different
                
                Most AI tutoring tools are **answer machines**. KnowLedge is a **verification system**.
                
                ### The Problem
                - Students copy AI-generated code without understanding
                - Traditional quizzes can be gamed with a second AI
                - Instructors have no way to detect this pattern
                
                ### The KnowLedge Solution
                - **Scout** logs every concept you borrow from AI
                - **Sage** forces you to explain it yourself before you own it
                - **Lens** checks handwritten work for logic gaps
                - **Integrity** fingerprints your session to catch repeat gaming
                - **Sync** gives instructors only concept-level aggregates (privacy-first)
                
                ### Why This Matters
                - Fixes a **real problem** in modern education
                - Uses **Gemma 4** for both extraction and dialogue
                - Runs **fully offline** for maximum privacy
                - Provides a **zero-setup** public demo (this Space)
                
                ### Key Metrics
                - **Debt Score**: How much of your work is still "borrowed"
                - **Spoof Attempts**: Session fingerprinting detects copy-paste patterns
                - **Integrity Signals**: Anti-gaming scoring from Lens
                
                This is what judges should look for. Not an answering machine. A **learning enforcer**.
                """
            )


# Launch the Gradio interface
if __name__ == "__main__":
    demo.launch()