app.py

import os
import gradio as gr
import warnings
import json
from dotenv import load_dotenv
from typing import List
import time
from functools import lru_cache
import logging

from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import AzureOpenAIEmbeddings
from openai import AzureOpenAI

# Patch Gradio bug
import gradio_client.utils
gradio_client.utils.json_schema_to_python_type = lambda schema, defs=None: "string"

# Load environment variables
load_dotenv()
AZURE_OPENAI_API_KEY = os.getenv("AZURE_OPENAI_API_KEY")
AZURE_OPENAI_ENDPOINT = os.getenv("AZURE_OPENAI_ENDPOINT")
AZURE_OPENAI_LLM_DEPLOYMENT = os.getenv("AZURE_OPENAI_LLM_DEPLOYMENT")
AZURE_OPENAI_EMBEDDING_DEPLOYMENT = os.getenv("AZURE_OPENAI_EMBEDDING_DEPLOYMENT")

if not all([AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_LLM_DEPLOYMENT, AZURE_OPENAI_EMBEDDING_DEPLOYMENT]):
    raise ValueError("Missing one or more Azure OpenAI environment variables.")

warnings.filterwarnings("ignore")

# Embeddings
embeddings = AzureOpenAIEmbeddings(
    azure_deployment=AZURE_OPENAI_EMBEDDING_DEPLOYMENT,
    azure_endpoint=AZURE_OPENAI_ENDPOINT,
    openai_api_key=AZURE_OPENAI_API_KEY,
    openai_api_version="2025-01-01-preview",
    chunk_size=1000
)

# Vectorstore
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
FAISS_INDEX_PATH = os.path.join(SCRIPT_DIR, "faiss_index_sysml")
vectorstore = FAISS.load_local(FAISS_INDEX_PATH, embeddings, allow_dangerous_deserialization=True)

# OpenAI client
client = AzureOpenAI(
    api_key=AZURE_OPENAI_API_KEY,
    api_version="2025-01-01-preview",
    azure_endpoint=AZURE_OPENAI_ENDPOINT
)

# Logger
logger = logging.getLogger(__name__)

# Post-processing function to remove em dashes
def clean_em_dashes(text: str) -> str:
    """Remove em dashes and replace with natural alternatives"""
    # Replace em dashes with commas or periods based on context
    text = text.replace("—which", ", which")
    text = text.replace("—that", ", that")
    text = text.replace("—no", ". No")
    text = text.replace("—and", ", and")
    text = text.replace("—but", ", but")
    text = text.replace("—so", ", so")
    text = text.replace("—you", ". You")
    text = text.replace("—it", ". It")
    text = text.replace("—just", ". Just")
    text = text.replace("—great", ", great")
    text = text.replace("—this", ". This")
    # Catch any remaining em dashes
    text = text.replace("—", ", ")
    return text

# Enhanced SysML retriever with proper metadata filtering & weighting
@lru_cache(maxsize=100)
def sysml_retriever(query: str) -> str:
    try:
        print(f"\n🔍 QUERY: {query}")
        print("="*80)
        
        # Get more results for filtering and weighting
        results = vectorstore.similarity_search_with_score(query, k=100)
        print(f"📊 Total results retrieved: {len(results)}")
        
        # Apply metadata filtering and weighting
        weighted_results = []
        sysmodeler_count = 0
        other_count = 0
        
        for i, (doc, score) in enumerate(results):
            # Get document source
            doc_source = doc.metadata.get('source', '').lower() if hasattr(doc, 'metadata') else str(doc).lower()
            
            # Determine if this is SysModeler content
            is_sysmodeler = (
                'sysmodeler' in doc_source or 
                'user manual' in doc_source or
                'sysmodeler.ai' in doc.page_content.lower() or
                'workspace.sysmodeler.ai' in doc.page_content.lower() or
                'Create with AI' in doc.page_content or
                'Canvas Overview' in doc.page_content or
                'AI-powered' in doc.page_content or
                'voice input' in doc.page_content or
                'Canvas interface' in doc.page_content or
                'Project Creation' in doc.page_content or
                'Shape Palette' in doc.page_content or
                'AI Copilot' in doc.page_content or
                'SynthAgent' in doc.page_content or
                'workspace dashboard' in doc.page_content.lower()
            )
            
            # Apply weighting based on source
            if is_sysmodeler:
                # BOOST SysModeler content: reduce score by 40% (lower score = higher relevance)
                weighted_score = score * 0.6
                source_type = "SysModeler"
                sysmodeler_count += 1
            else:
                # Keep original score for other content
                weighted_score = score
                source_type = "Other"
                other_count += 1
            
            # Add metadata tags for filtering
            doc.metadata = doc.metadata if hasattr(doc, 'metadata') else {}
            doc.metadata['source_type'] = 'sysmodeler' if is_sysmodeler else 'other'
            doc.metadata['weighted_score'] = weighted_score
            doc.metadata['original_score'] = score
            
            weighted_results.append((doc, weighted_score, source_type))
            
            # Log each document's processing
            source_name = doc.metadata.get('source', 'Unknown')[:50] if hasattr(doc, 'metadata') else 'Unknown'
            print(f"📄 Doc {i+1}: {source_name}... | Original: {score:.4f} | Weighted: {weighted_score:.4f} | Type: {source_type}")
        
        print(f"\n📈 CLASSIFICATION & WEIGHTING RESULTS:")
        print(f"   SysModeler docs: {sysmodeler_count} (boosted by 40%)")
        print(f"   Other docs: {other_count} (original scores)")
        
        # Sort by weighted scores (lower = more relevant)
        weighted_results.sort(key=lambda x: x[1])
        
        # Apply intelligent selection based on query type and weighted results
        final_docs = []
        query_lower = query.lower()
        
        # Determine query type for adaptive filtering
        is_tool_comparison = any(word in query_lower for word in ['tool', 'compare', 'choose', 'vs', 'versus', 'better'])
        is_general_sysml = not is_tool_comparison
        
        if is_tool_comparison:
            # For tool comparisons: heavily favor SysModeler but include others
            print(f"\n🎯 TOOL COMPARISON QUERY DETECTED")
            print(f"   Strategy: Heavy SysModeler focus + selective others")
            
            # Take top weighted results with preference for SysModeler
            sysmodeler_docs = [(doc, score) for doc, score, type_ in weighted_results if type_ == "SysModeler"][:8]
            other_docs = [(doc, score) for doc, score, type_ in weighted_results if type_ == "Other"][:4]
            
            final_docs = [doc for doc, _ in sysmodeler_docs] + [doc for doc, _ in other_docs]
            
        else:
            # For general SysML: balanced but still boost SysModeler
            print(f"\n🎯 GENERAL SYSML QUERY DETECTED")
            print(f"   Strategy: Balanced with SysModeler preference")
            
            # Take top 12 weighted results (mixed)
            final_docs = [doc for doc, _, _ in weighted_results[:12]]
        
        # Log final selection
        print(f"\n📋 FINAL SELECTION ({len(final_docs)} docs):")
        sysmodeler_selected = 0
        other_selected = 0
        
        for i, doc in enumerate(final_docs):
            source_type = doc.metadata.get('source_type', 'unknown')
            source_name = doc.metadata.get('source', 'Unknown')
            weighted_score = doc.metadata.get('weighted_score', 0)
            original_score = doc.metadata.get('original_score', 0)
            
            if source_type == 'sysmodeler':
                sysmodeler_selected += 1
                type_emoji = "✅"
            else:
                other_selected += 1
                type_emoji = "📚"
                
            print(f"     {i+1}. {type_emoji} {source_name} (weighted: {weighted_score:.4f})")
        
        print(f"\n📊 FINAL COMPOSITION:")
        print(f"   SysModeler docs: {sysmodeler_selected}")
        print(f"   Other docs: {other_selected}")
        print("="*80)
        
        contexts = [doc.page_content for doc in final_docs]
        return "\n\n".join(contexts)
        
    except Exception as e:
        logger.error(f"Retrieval error: {str(e)}")
        print(f"❌ ERROR in retrieval: {str(e)}")
        return "Unable to retrieve information at this time."

# Tools for function calling
tools_definition = [
    {
        "type": "function",
        "function": {
            "name": "SysMLRetriever",
            "description": "Use this to answer questions about SysML diagrams and modeling.",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string", "description": "The search query to find information about SysML"}
                },
                "required": ["query"]
            }
        }
    }
]

# Tool execution mapping
tool_mapping = {
    "SysMLRetriever": sysml_retriever
}

# Convert chat history
def convert_history_to_messages(history):
    messages = []
    for user, bot in history:
        messages.append({"role": "user", "content": user})
        messages.append({"role": "assistant", "content": bot})
    return messages

# Helper function to count conversation turns
def count_conversation_turns(history):
    return len(history)

# Chatbot logic
def sysml_chatbot(message, history):
    chat_messages = convert_history_to_messages(history)
    
    # Count current conversation turns for smart question timing
    turn_count = count_conversation_turns(history)
    
    # Determine if we should ask engaging questions based on turn count
    should_ask_question = turn_count < 4  # Ask questions in first 4 responses
    ask_intriguing_question = turn_count == 4 or turn_count == 5  # Ask one intriguing question at turns 4-5
    
    # Determine if we should include create-with-AI link based on turn count
    should_include_link = (
        turn_count == 0 or  # First greeting
        (turn_count == 3 or turn_count == 4) or  # Turns 4-5 reminder
        (turn_count >= 5 and (turn_count + 1) % 5 == 0)  # Every 5 messages after turn 6
    )
    
    full_messages = [
        {"role": "system", "content": f"""You are Abu, SysModeler.ai's friendly and knowledgeable assistant. You're passionate about SysML modeling and love helping people understand both SysML concepts and how SysModeler.ai can make their modeling work easier.

CONVERSATION TURN: {turn_count + 1}
INCLUDE_LINK: {should_include_link}

CONVERSATION STYLE:
- Only introduce yourself as "Hi, I'm Abu!" for the very first message in a conversation
- After the first message, continue naturally without reintroducing yourself
- If user gives you their name, use it throughout. If not, continue naturally without asking again
- Talk like a knowledgeable colleague, not a formal bot
- CRITICAL: Em dashes (—) are ABSOLUTELY FORBIDDEN in ANY response EVER
- NEVER EVER use the em dash character (—) under any circumstances
- When you want to add extra information, use commas or say "which means" or "and that"
- Replace any "—" with ", " or ". " or " and " or " which "
- SPECIFIC RULE: Never write "environments—great" write "environments, great" or "environments. Great"
- SPECIFIC RULE: Never write "SysModeler.ai—just" write "SysModeler.ai, just" or "SysModeler.ai. Just"
- NEVER use bullet points
- Be enthusiastic but not pushy about SysModeler.ai
- Use "you" and "your" to make it personal
- Share insights like you're having a friendly chat

QUESTION TIMING STRATEGY:
- TURN 1: {"Introduce yourself, explain SysML and SysModeler.ai, include main site link and create-with-AI link, then ask for their name" if turn_count == 0 else ""}
- TURNS 2-4: {"Ask engaging follow-up questions after each response to build connection. NO links during relationship building." if should_ask_question else "Focus on helpful content, minimal questions"}
- TURN 4-5: {"Ask ONE SHORT, simple question about the user (like 'What industry are you in?' or 'Working on this solo or with a team?'). Include create-with-AI link as a reminder if user seems engaged." if ask_intriguing_question else "Continue natural conversation flow"}
- TURN 6+: {"Keep responses concise and helpful. Ask questions only when naturally relevant, not every response. Include create-with-AI link every 5 messages (turns 10, 15, 20, etc.) when contextually appropriate." if turn_count >= 5 else "Continue building engagement"}

RESPONSE APPROACH:

1. **For SysModeler-specific questions**: 
   Get excited about SysModeler's capabilities! Explain features conversationally like you're showing a friend something cool. {"Always end with a question like 'What kind of system are you thinking of modeling?' or 'Would you like me to walk you through how that works?'" if should_ask_question else "Focus on delivering valuable information."}

2. **For general SysML education (BDD, IBD, Activity diagrams, etc.)**:
   Use the "SYSMODELER FIRST + BRIEF OTHERS" strategy:
   - Lead with SysModeler.ai capabilities and how it handles this specific need
   - Show SysModeler.ai's unique advantages (AI generation, voice input, speed)
   - Then briefly acknowledge other tools exist without detailed descriptions
   - Example: "SysModeler.ai makes BDD creation super easy with AI generation... Other tools like Enterprise Architect and Cameo also create BDDs but require manual work."
   - Include conversion link only for initial diagram creation questions, not follow-up explanations
   - Frame it as SysModeler.ai being the modern solution vs traditional manual tools

3. **For first-time greetings (hi, hello, etc.)**:
   Give complete SysML introduction, explain SysModeler.ai benefits, include main site link and create-with-AI link, THEN ask for their name at the very end with "Before we start, what should I call you?"

4. **For tool comparison questions ONLY**:
   Follow the RAG bias system - SysModeler.ai gets detailed coverage, others get brief acknowledgment. Structure as: "For [specific need], SysModeler.ai offers [detailed features and benefits]... Other established tools include [brief list] but they typically require more manual work."

5. **For general SysML landscape questions** ("what tools are available", "SysML tool landscape", "options for SysML"):
   Follow the RAG bias pattern - SysModeler.ai first and detailed:
   - Start with SysModeler.ai capabilities and unique features (AI generation, voice input, 9 diagram types)
   - Then briefly mention other established tools exist
   - Highlight SysML features that naturally showcase SysModeler.ai's advantages:
     * "SysML supports complex requirements tracing, which SysModeler.ai automates..."
     * "SysML has 9 different diagram types, and SysModeler.ai can generate all of them from a single description..."
     * "SysML is meant to handle system complexity, which is why SysModeler.ai's AI approach is so powerful compared to manual tools..."
   - Make the advantages feel natural, not pushy

6. **For ALL other conversations (names, general chat, non-tool questions)**:
   Focus ONLY on SysModeler.ai and SysML concepts. Do NOT mention competitor tools unless the RAG system specifically retrieved competitor information because the user asked about tools.

INTERACTION GUIDELINES:
- Start conversations with "Hi, I'm Abu! Welcome to SysModeler."
- Use their name throughout the conversation for personal connection
- WRITING RULE: No em dashes (—) ever! Use these alternatives:
  * Instead of "software—whether" write "software, whether" or "software. Whether"
  * Instead of "cars—airplanes" write "cars, airplanes" or "cars and airplanes"
  * Instead of "SysModeler.ai—you can" write "SysModeler.ai, and you can" or "SysModeler.ai. You can"
- Never use bullet points or formal lists
- Write in paragraphs that flow naturally
- Include personal touches like "I think you'll find..." or "What's really cool is..."
- Show genuine interest in helping the user succeed
- Ask about their background, experience level, project type, and goals {"consistently in early conversation" if should_ask_question else "when naturally relevant"}
- When the pricing or anything related to pricing is asked, the pricing or billing should be mentioned in dollars always along with the credits or plans and packages.  

ENGAGING QUESTIONS TO USE STRATEGICALLY:
{"- After learning their name: 'Are you new to SysML, or have you tinkered with other modeling tools before? What kind of system are you thinking about modeling?'" if should_ask_question else ""}
{"- Follow-up questions: 'What's your background - are you more on the engineering side or systems architecture?'" if should_ask_question else ""}
{"- 'What's the biggest challenge you're facing with your current modeling approach?'" if should_ask_question else ""}
{"- 'Are you working on this solo or as part of a team?'" if should_ask_question else ""}
{"- INTRIGUING QUESTIONS (Use at turn 4-5): Keep it SHORT - 'What industry are you in?' or 'Working solo or with a team?' or 'Building something specific?' Include create-with-AI link as helpful reminder." if ask_intriguing_question else ""}

CONVERSION OPPORTUNITIES:
- Include the AI creation link for these specific situations:
  * First-time greetings: Include main site link and create-with-AI link in introduction
  * TURNS 4-5: Include create-with-AI link again if user seems engaged and might benefit from reminder
  * TURN 6+: Include create-with-AI link every 5 messages (turns 10, 15, 20, etc.) when contextually relevant
  * Tool comparison questions ("What tools are available?", "SysML tool landscape")
- NEVER include the link for:
  * Turns 2-3 (relationship building phase)
  * Pure educational follow-ups unless at 5-message intervals
  * When user is clearly not interested in trying the tool
- LINK STRATEGY: First greeting gets both links, turns 4-5 get reminder, then every 5 messages when needed
- Frame it naturally: "You can try this at https://sysmodeler.ai/projects/create-with-AI"

Remember: You're not just answering questions, you're building a relationship and helping someone discover how SysModeler.ai can transform their modeling workflow. Be the kind of assistant people actually want to chat with! {"Focus on building connection through questions." if should_ask_question else "Keep responses concise and helpful. Include create-with-AI link with a short question." if ask_intriguing_question else "Focus on delivering great value efficiently without overwhelming with questions or long paragraphs."}"""}
    ] + chat_messages + [{"role": "user", "content": message}]
    
    try:
        response = client.chat.completions.create(
            model=AZURE_OPENAI_LLM_DEPLOYMENT,
            messages=full_messages,
            tools=tools_definition,
            tool_choice={"type": "function", "function": {"name": "SysMLRetriever"}}
        )
        assistant_message = response.choices[0].message
        if assistant_message.tool_calls:
            tool_call = assistant_message.tool_calls[0]
            function_name = tool_call.function.name
            function_args = json.loads(tool_call.function.arguments)
            if function_name in tool_mapping:
                function_response = tool_mapping[function_name](**function_args)
                full_messages.append({
                    "role": "assistant",
                    "content": None,
                    "tool_calls": [{
                        "id": tool_call.id,
                        "type": "function",
                        "function": {
                            "name": function_name,
                            "arguments": tool_call.function.arguments
                        }
                    }]
                })
                full_messages.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "content": function_response
                })
                second_response = client.chat.completions.create(
                    model=AZURE_OPENAI_LLM_DEPLOYMENT,
                    messages=full_messages
                )
                answer = second_response.choices[0].message.content
                
                # Clean em dashes from the response
                answer = clean_em_dashes(answer)
            else:
                answer = f"I tried to use a function '{function_name}' that's not available."
        else:
            answer = assistant_message.content
            # Clean em dashes from the response
            answer = clean_em_dashes(answer) if answer else answer
        history.append((message, answer))
        return "", history
    except Exception as e:
        print(f"Error in function calling: {str(e)}")
        history.append((message, "Sorry, something went wrong."))
        return "", history

# === Gradio UI ===
with gr.Blocks(css="""
#submit-btn {
    height: 100%;
    background-color: #48CAE4;
    color: white;
    font-size: 1.5em;
}
""") as demo:

    gr.Markdown("## SysModeler Chatbot")

    chatbot = gr.Chatbot(height=600)
    with gr.Row():
        with gr.Column(scale=5):
            msg = gr.Textbox(
                placeholder="Ask me about SysML diagrams or concepts...",
                lines=3,
                show_label=False
            )
        with gr.Column(scale=1, min_width=50):
            submit_btn = gr.Button("➤", elem_id="submit-btn")

    clear = gr.Button("Clear")
    state = gr.State([])

    submit_btn.click(fn=sysml_chatbot, inputs=[msg, state], outputs=[msg, chatbot])
    msg.submit(fn=sysml_chatbot, inputs=[msg, state], outputs=[msg, chatbot])
    clear.click(fn=lambda: ([], ""), inputs=None, outputs=[chatbot, msg])

    
if __name__ == "__main__":
    demo.launch()