app_backup.py

"""
Krish Mind Local Server (GGUF)
==============================
Works with index_local.html
Features: GGUF Model + Web Search + Image Generation + RAG + File Upload
"""

import os
import sys
import urllib.parse
import pickle
import tempfile
from datetime import datetime

print("=" * 60)
print("🧠 Krish Mind Local Server (GGUF)")
print("=" * 60)

# --- Core dependencies ---
try:
    from llama_cpp import Llama
    print("✅ llama-cpp-python")
except ImportError:
    print("❌ Run: pip install llama-cpp-python")
    sys.exit(1)

try:
    from fastapi import FastAPI, UploadFile, File
    from fastapi.middleware.cors import CORSMiddleware
    from pydantic import BaseModel
    import uvicorn
    print("✅ fastapi + uvicorn")
except ImportError:
    print("❌ Run: pip install fastapi uvicorn python-multipart")
    sys.exit(1)

# --- Config ---
GGUF_PATH = "d:/Krish Mind/gguf/krish-mind-standalone-Q4.gguf"
EMBEDDINGS_FILE = "../data/krce_embeddings.pkl"
DATA_FILE = "../data/krce_college_data.jsonl"

# --- Load GGUF Model ---
print(f"\n⏳ Loading GGUF model...")
try:
    model = Llama(
        model_path=GGUF_PATH,
        n_ctx=4096,
        n_gpu_layers=0,
        verbose=False
    )
    print("✅ Model loaded!")
except Exception as e:
    print(f"❌ Model error: {e}")
    sys.exit(1)

# --- DuckDuckGo Web Search ---
print("\n📦 Loading optional features...")
ddgs = None
try:
    import warnings
    warnings.filterwarnings("ignore")
    from duckduckgo_search import DDGS
    ddgs = DDGS()
    print("✅ DuckDuckGo web search")
except Exception as e:
    print(f"⚠️ Web search disabled: {e}")

# --- RAG SETUP (Load Pre-computed Embeddings) ---
print("📚 Loading Knowledge Base...")
knowledge_base = []
doc_embeddings = None
rag_model = None

# Try to load pre-computed embeddings first
if os.path.exists(EMBEDDINGS_FILE):
    try:
        import numpy as np
        from sentence_transformers import SentenceTransformer
        
        print(f"📂 Loading pre-computed embeddings from {EMBEDDINGS_FILE}...")
        with open(EMBEDDINGS_FILE, 'rb') as f:
            data = pickle.load(f)
        
        knowledge_base = data['knowledge_base']
        doc_embeddings = data['embeddings']
        
        # Load the model for query encoding (needed for search)
        rag_model = SentenceTransformer(data.get('model_name', 'all-MiniLM-L6-v2'))
        print(f"✅ Embeddings loaded! ({len(knowledge_base)} facts)")
        
    except Exception as e:
        print(f"⚠️ Could not load embeddings: {e}")
        print("   Falling back to live embedding...")

# Fallback: compute embeddings if pkl not found
if doc_embeddings is None and os.path.exists(DATA_FILE):
    try:
        from sentence_transformers import SentenceTransformer
        import numpy as np
        import json
        
        rag_model = SentenceTransformer('all-MiniLM-L6-v2')
        print("✅ Embedding model loaded (SentenceTransformer)")
        
        with open(DATA_FILE, 'r') as f:
            for line in f:
                if line.strip():
                    try:
                        knowledge_base.append(json.loads(line))
                    except:
                        pass
        
        if knowledge_base:
            docs = [f"{k['instruction']} {k['output']}" for k in knowledge_base]
            doc_embeddings = rag_model.encode(docs)
            print(f"✅ Embeddings computed! ({len(knowledge_base)} facts)")
            print("   ⚠️ Run 'python scripts/build_embeddings.py' for faster startup!")
            
    except Exception as e:
        print(f"❌ RAG disabled: {e}")
        rag_model = None
else:
    if doc_embeddings is None:
        print("⚠️ Data file not found! RAG disabled.")

# Initialize Cross-Encoder for re-ranking
cross_encoder = None
try:
    from sentence_transformers import CrossEncoder
    cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
    print("✅ Cross-Encoder loaded for re-ranking")
except Exception as e:
    print(f"⚠️ Cross-Encoder not available: {e}")

# ============================================
# FILE UPLOAD RAG: Session-based document analysis
# ============================================
# Store uploaded file embeddings per session (in-memory)
session_file_data = {}

def extract_text_from_file(file_path: str, filename: str) -> str:
    """Extract text from uploaded file"""
    text = ""
    ext = filename.lower().split('.')[-1]
    
    try:
        if ext == 'txt':
            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
                text = f.read()
        
        elif ext == 'pdf':
            try:
                import PyPDF2
                with open(file_path, 'rb') as f:
                    reader = PyPDF2.PdfReader(f)
                    for page in reader.pages:
                        text += page.extract_text() + "\n"
            except ImportError:
                try:
                    import fitz  # PyMuPDF
                    doc = fitz.open(file_path)
                    for page in doc:
                        text += page.get_text() + "\n"
                    doc.close()
                except ImportError:
                    return "Error: Install PyPDF2 or PyMuPDF to read PDFs"
        
        elif ext in ['doc', 'docx']:
            try:
                import docx
                doc = docx.Document(file_path)
                text = "\n".join([para.text for para in doc.paragraphs])
            except ImportError:
                return "Error: Install python-docx to read Word files"
        
        elif ext in ['md', 'json', 'csv']:
            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
                text = f.read()
        
        else:
            # Try reading as plain text
            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
                text = f.read()
    
    except Exception as e:
        return f"Error reading file: {e}"
    
    return text.strip()

def chunk_text(text: str, chunk_size: int = 500, overlap: int = 50) -> list:
    """Split text into chunks for embedding"""
    words = text.split()
    chunks = []
    
    for i in range(0, len(words), chunk_size - overlap):
        chunk = " ".join(words[i:i + chunk_size])
        if chunk.strip():
            chunks.append(chunk)
    
    return chunks

# ============================================
# ADVANCED RAG: Query Expansion + Hybrid Search
# ============================================

ABBREVIATIONS = {
    "aids": "AI&DS Artificial Intelligence and Data Science",
    "ai&ds": "AI&DS Artificial Intelligence and Data Science",
    "aid": "AI&DS Artificial Intelligence and Data Science",
    "aiml": "AI&ML Artificial Intelligence and Machine Learning",
    "ai&ml": "AI&ML Artificial Intelligence and Machine Learning",
    "cse": "Computer Science Engineering CSE",
    "ece": "Electronics Communication Engineering ECE",
    "eee": "Electrical Electronics Engineering EEE",
    "mech": "Mechanical Engineering",
    "it": "Information Technology IT",
    "hod": "Head of Department HOD",
    "mam": "madam professor female faculty",
    "sir": "male professor faculty",
    "staffs": "staff faculty members",
    "sase": "Sasikumar Sasidevi",
    "krce": "K. Ramakrishnan College of Engineering",
}

def expand_query(query):
    """Expand abbreviations and add synonyms for better matching"""
    expanded = query.lower()
    for abbr, full in ABBREVIATIONS.items():
        if abbr in expanded.split():
            expanded = expanded + " " + full
    return expanded

def search_krce(query, threshold=0.25):
    """Advanced RAG: Query Expansion + Vector Search + Cross-Encoder Re-ranking"""
    if not rag_model or doc_embeddings is None:
        return ""
    
    try:
        expanded_query = expand_query(query)
        
        print(f"\n📊 RAG Search")
        print(f"  Original: '{query}'")
        print(f"  Expanded: '{expanded_query}'")
        print("-" * 50)
        
        from sklearn.metrics.pairwise import cosine_similarity
        q_emb = rag_model.encode([expanded_query])
        vector_scores = cosine_similarity(q_emb, doc_embeddings).flatten()
        
        top_indices = vector_scores.argsort()[-10:][::-1]
        top_candidates = [(idx, vector_scores[idx]) for idx in top_indices]

        print("Vector Search Results:")
        for i, (idx, v) in enumerate(top_candidates[:5]):
            instruction = knowledge_base[idx]['instruction'][:35]
            print(f"  #{i+1} V:{v:.3f} | {instruction}...")
        
        if cross_encoder:
            pairs = [[query, f"{knowledge_base[idx]['instruction']} {knowledge_base[idx]['output']}"] 
                     for idx, _ in top_candidates]
            ce_scores = cross_encoder.predict(pairs)
            
            final_ranking = sorted(zip(top_candidates, ce_scores), key=lambda x: x[1], reverse=True)
            
            print("Cross-Encoder Re-ranking:")
            for i, ((idx, v), ce) in enumerate(final_ranking[:3]):
                instruction = knowledge_base[idx]['instruction'][:35]
                print(f"  #{i+1} CE:{ce:.3f} | {instruction}...")
            print("-" * 50)
            
            final_context = []
            print(f"✅ RAG Retrieval (Top 5):")
            for i, ((idx, v), ce) in enumerate(final_ranking[:5]):
                if ce > -6.0:
                    content = knowledge_base[idx]['output']
                    final_context.append(content)
                    print(f"  Took #{i+1}: {knowledge_base[idx]['instruction'][:30]}...")
            
            if final_context:
                return "\n\n".join(final_context)
        else:
            final_context = []
            for i, (idx, score) in enumerate(top_candidates[:5]):
                if score > threshold:
                    final_context.append(knowledge_base[idx]['output'])
            
            if final_context:
                return "\n\n".join(final_context)
        
        print("❌ No confident match found")
        return ""
        
    except Exception as e:
        print(f"RAG Error: {e}")
        return ""

def search_file_context(query: str, session_id: str) -> str:
    """Search uploaded file for relevant context"""
    if session_id not in session_file_data or not rag_model:
        print(f"⚠️ File context not found: session_id={session_id}, in_session={session_id in session_file_data}")
        return ""
    
    # Maximum characters to return (prevent context overflow)
    MAX_CONTEXT_CHARS = 2000
    
    try:
        file_data = session_file_data[session_id]
        chunks = file_data['chunks']
        embeddings = file_data['embeddings']
        filename = file_data.get('filename', 'uploaded file')
        
        # Detect general queries about the file (summarize, what's in the file, etc.)
        general_triggers = ['summarize', 'summary', 'what is in', "what's in", 'tell me about', 
                           'describe', 'overview', 'main points', 'key points', 'the file', 
                           'the document', 'uploaded', 'attached', 'read the']
        is_general_query = any(t in query.lower() for t in general_triggers)
        
        if is_general_query:
            # For general queries, return content up to limit
            print(f"📄 General file query detected - returning limited content")
            all_content = ""
            for chunk in chunks:
                if len(all_content) + len(chunk) > MAX_CONTEXT_CHARS:
                    break
                all_content += chunk + "\n\n"
            
            if len(all_content) < sum(len(c) for c in chunks):
                all_content += f"\n[...content truncated, showing {len(all_content)} of {sum(len(c) for c in chunks)} chars]"
            
            return f"[Content from {filename}]:\n\n{all_content.strip()}"
        
        # For specific queries, use semantic search
        from sklearn.metrics.pairwise import cosine_similarity
        q_emb = rag_model.encode([query])
        scores = cosine_similarity(q_emb, embeddings).flatten()
        
        # Get top 3 most relevant chunks
        top_indices = scores.argsort()[-3:][::-1]
        context_parts = []
        total_chars = 0
        
        for idx in top_indices:
            if scores[idx] > 0.15 and total_chars < MAX_CONTEXT_CHARS:
                chunk = chunks[idx]
                if total_chars + len(chunk) > MAX_CONTEXT_CHARS:
                    # Truncate to fit
                    remaining = MAX_CONTEXT_CHARS - total_chars
                    chunk = chunk[:remaining] + "..."
                context_parts.append(chunk)
                total_chars += len(chunk)
        
        if context_parts:
            print(f"📄 File context found ({len(context_parts)} chunks, {total_chars} chars)")
            return f"[Content from {filename}]:\n\n" + "\n\n".join(context_parts)
        
        # Fallback: if no good matches, return first chunk truncated
        print(f"📄 Low confidence match - returning truncated first chunk")
        first_chunk = chunks[0][:MAX_CONTEXT_CHARS] if chunks else ""
        return f"[Content from {filename}]:\n\n{first_chunk}"
        
    except Exception as e:
        print(f"File search error: {e}")
        return ""

def search_web(query):
    if not ddgs:
        return ""
    try:
        results = ddgs.text(query, max_results=3)
        if not results:
            return ""
        return "\n\n".join([f"**{r['title']}**\n{r['body']}" for r in results])
    except:
        return ""

# --- FastAPI ---
app = FastAPI()
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])

class ChatRequest(BaseModel):
    message: str
    max_tokens: int = 512
    temperature: float = 0.7
    summary: str = ""
    history: list = []
    session_id: str = ""  # For file upload sessions

class SummarizeRequest(BaseModel):
    messages: list

@app.get("/")
async def root():
    return {"name": "Krish Mind", "status": "online", "rag": rag_model is not None, "web": ddgs is not None}

@app.post("/upload")
async def upload_file(file: UploadFile = File(...), session_id: str = "default"):
    """Upload a file for RAG analysis"""
    if not rag_model:
        return {"success": False, "error": "Embedding model not available"}
    
    try:
        # Save uploaded file temporarily
        suffix = '.' + file.filename.split('.')[-1] if '.' in file.filename else '.txt'
        with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
            content = await file.read()
            tmp.write(content)
            tmp_path = tmp.name
        
        # Extract text
        print(f"📤 Processing uploaded file: {file.filename}")
        text = extract_text_from_file(tmp_path, file.filename)
        
        # Clean up temp file
        os.unlink(tmp_path)
        
        if text.startswith("Error"):
            return {"success": False, "error": text}
        
        if not text.strip():
            return {"success": False, "error": "Could not extract text from file"}
        
        # Chunk text
        chunks = chunk_text(text)
        if not chunks:
            return {"success": False, "error": "File too small or empty"}
        
        # Create embeddings
        print(f"🔄 Creating embeddings for {len(chunks)} chunks...")
        embeddings = rag_model.encode(chunks)
        
        # Store in session
        session_file_data[session_id] = {
            "filename": file.filename,
            "chunks": chunks,
            "embeddings": embeddings,
            "full_text": text[:2000]  # First 2000 chars for context
        }
        
        print(f"✅ File processed: {len(chunks)} chunks, {len(text)} chars")
        
        return {
            "success": True,
            "filename": file.filename,
            "chunks": len(chunks),
            "chars": len(text),
            "preview": text[:200] + "..." if len(text) > 200 else text
        }
        
    except Exception as e:
        print(f"❌ Upload error: {e}")
        return {"success": False, "error": str(e)}

@app.delete("/upload/{session_id}")
async def clear_file(session_id: str):
    """Clear uploaded file from session"""
    if session_id in session_file_data:
        del session_file_data[session_id]
        return {"success": True, "message": "File cleared"}
    return {"success": False, "message": "No file found for session"}

@app.post("/summarize")
async def summarize(request: SummarizeRequest):
    """Summarize older messages to compress context"""
    try:
        messages_text = ""
        for msg in request.messages:
            role = msg.get("role", "user")
            content = msg.get("content", "")
            messages_text += f"{role.capitalize()}: {content}\n"
        
        summary_prompt = f"""<|start_header_id|>system<|end_header_id|>

You are a conversation summarizer. Condense the following conversation into a brief summary (2-3 sentences max) that captures the key topics and context. Focus on what was discussed, not exact words.<|eot_id|><|start_header_id|>user<|end_header_id|>

Summarize this conversation:
{messages_text}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Summary: """
        
        output = model(summary_prompt, max_tokens=150, temperature=0.3, stop=["<|eot_id|>"], echo=False)
        summary = output["choices"][0]["text"].strip()
        print(f"📝 Summarized {len(request.messages)} messages: {summary[:50]}...")
        return {"summary": summary}
    except Exception as e:
        print(f"❌ Summarization error: {e}")
        return {"summary": "", "error": str(e)}


@app.post("/chat")
async def chat(request: ChatRequest):
    user_input = request.message
    session_id = request.session_id or "default"
    
    # Image generation
    img_triggers = ["generate image", "create image", "draw", "imagine"]
    if any(t in user_input.lower() for t in img_triggers):
        prompt = user_input
        for t in img_triggers:
            prompt = prompt.lower().replace(t, "")
        prompt = prompt.strip()
        if prompt:
            url = f"https://image.pollinations.ai/prompt/{urllib.parse.quote(prompt)}"
            return {"response": f"Here's your image of **{prompt}**:\n\n![{prompt}]({url})"}
    
    # RAG Search (college knowledge base)
    rag_context = ""
    if rag_model:
        rag_context = search_krce(user_input)
        if rag_context:
            print(f"🧠 RAG Context found: {rag_context[:50]}...")
    
    # File context (uploaded document)
    file_context = ""
    if session_id in session_file_data:
        file_context = search_file_context(user_input, session_id)
        if file_context:
            print(f"📄 File Context found: {file_context[:50]}...")

    # Web search
    web_context = ""
    search_triggers = ["search", "find", "latest", "news", "who is", "what is", "when", "where", "how"]
    if ddgs and any(t in user_input.lower() for t in search_triggers):
        if len(user_input.split()) > 2:
            print(f"🔎 Searching web...")
            web_context = search_web(user_input)
    
    # Build prompt
    now = datetime.now().strftime("%A, %B %d, %Y %I:%M %p")
    
    sys_prompt = f"""You are Krish Mind, a helpful AI assistant created by Krish CS. Current time: {now}

IMPORTANT STRICT RULES:
1. IDENTITY: You were created by Krish CS. Do NOT claim to be created by anyone mentioned in the context (like faculty, HODs, or staff). If the context mentions a name, that person is a subject of the data, NOT your creator.
2. CONTEXT USAGE: Use the provided context to answer questions. If the context contains a list (e.g., faculty names), make sure to include ALL items found in the context chunks.
3. FORMATTING: Use Markdown. For letters, use **bold** for headers (e.g., **Subject:**) and use DOUBLE LINE BREAKS between sections (Place, Date, From, To, Subject, Body) to create clear distinct paragraphs.
4. AMBIGUITY: 'AID' or 'AIDS' in this context ALWAYS refers to 'Artificial Intelligence and Data Science', NEVER 'Aerospace' or 'Disease'.
5. ACCURACY: If the context contains a name like 'Mrs. C. Rani', she is a faculty member. Do NOT say "I was created by Mrs. C. Rani".
"""
    
    if file_context:
        sys_prompt += f"\n\nUploaded Document Context:\n{file_context}"
    
    if rag_context:
        sys_prompt += f"\n\nKnowledge Base Context:\n{rag_context}"
        
    if web_context:
        sys_prompt += f"\n\nWeb Results:\n{web_context}"
    
    if request.summary:
        sys_prompt += f"\n\nPrevious conversation summary:\n{request.summary}"
    
    # Build history context
    history_context = ""
    if request.history:
        for msg in request.history[-6:]:
            role = msg.get("role", "user")
            content = msg.get("content", "")
            if role == "user":
                history_context += f"<|start_header_id|>user<|end_header_id|>\n\n{content}<|eot_id|>"
            else:
                history_context += f"<|start_header_id|>assistant<|end_header_id|>\n\n{content}<|eot_id|>"
    
    # Build full prompt
    if history_context:
        full_prompt = f"""<|start_header_id|>system<|end_header_id|>

{sys_prompt}<|eot_id|>{history_context}<|start_header_id|>user<|end_header_id|>

{user_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
    else:
        full_prompt = f"""<|start_header_id|>system<|end_header_id|>

{sys_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{user_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
    
    try:
        print(f"💬 Generating response...")
        output = model(full_prompt, max_tokens=request.max_tokens, temperature=request.temperature, stop=["<|eot_id|>"], echo=False)
        response = output["choices"][0]["text"].strip()
        print(f"✅ Done")
        return {"response": response}
    except Exception as e:
        return {"response": f"Error: {e}"}

if __name__ == "__main__":
    print("\n" + "=" * 60)
    print("🚀 Server running at: http://127.0.0.1:8000")
    print("📱 Open index_local.html in your browser")
    print("=" * 60 + "\n")
    uvicorn.run(app, host="0.0.0.0", port=8000)