app.py

from flask import Flask, render_template, request, jsonify
import faiss
import numpy as np
import json
from sentence_transformers import SentenceTransformer
from langchain.prompts import PromptTemplate
from langchain_groq import ChatGroq
import re
from langchain_openai import AzureChatOpenAI
import os
from dotenv import load_dotenv
load_dotenv()

app = Flask(__name__)

# Load Model and FAISS Index


AzureOpenAiEndpoint= os.getenv("AZURE_OPENAI_ENDPOINT")
AzureOpenAiKey= os.getenv("AZURE_OPENAI_KEY")
OpenaiApiVersion= os.getenv("OPENAI_API_VERSION")
DeploymentName= os.getenv("DEPLOYMENT_NAME")

model = SentenceTransformer('./sentence-transformers_all-MiniLM-L6-v2')
index = faiss.read_index("faiss_index.bin")
# groq_api_key = os.getenv('GROQ_API_KEY')
# model_name = "llama-3.3-70b-versatile"

llm = AzureChatOpenAI(
    openai_api_version= OpenaiApiVersion,  # Change based on your API version
    deployment_name = DeploymentName,  # The model deployment name in Azure
    openai_api_key= AzureOpenAiKey,  # Use the environment variable
    azure_endpoint=os.environ["AZURE_OPENAI_ENDPOINT"],  # Use the environment variable
)
# llm = ChatGroq(
#             temperature=0, 
#             groq_api_key=groq_api_key, 
#             model_name=model_name
#         )

with open("metadata.json") as f:
    metadata = json.load(f)


def categorize_query(query):
    """
    Categorizes user queries into different types (greetings, small talk, unrelated, etc.).
    """
    query = query.lower().strip()

    # Greetings
    greeting_patterns = [
        r"\bhello\b", r"\bhi\b", r"\bhey\b", r"\bhola\b", r"\bgreetings\b",
        r"\bwhat('s| is) up\b", r"\bhowdy\b", r"\bhiya\b", r"\byo\b",
        r"\bgood (morning|afternoon|evening|day|night)\b",
        r"\bhow (are|r) you\b", r"\bhow's it going\b", r"\bhow have you been\b",
        r"\bhope you are (doing )?(well|good|fine)\b", r"\bnice to meet you\b",
        r"\bpleased to meet you\b"
    ]

    # Thank-you messages
    thank_you_patterns = [
        r"\bthank(s| you)\b", r"\bthanks a lot\b", r"\bthanks so much\b",
        r"\bthank you very much\b", r"\bappreciate it\b", r"\bmuch obliged\b",
        r"\bgrateful\b", r"\bcheers\b"
    ]

    # Small talk
    small_talk_patterns = [
        r"\bhow (are|r) you\b", r"\bhow's your day\b", r"\bwhat's up\b",
        r"\bhow's it going\b", r"\bhow have you been\b", r"\bhope you are well\b"
    ]

    # Unrelated topics
    unrelated_patterns = [
        r"\btell me a joke\b", r"\bwho won\b", r"\bwhat is ai\b", r"\bexplain blockchain\b"
    ]


    # Goodbye messages
    goodbye_patterns = [
        r"\bbye\b", r"\bgoodbye\b", r"\bsee you\b", r"\bhave a nice day\b"
    ]

    # Rude or inappropriate messages
    rude_patterns = [
        r"\bstupid\b", r"\bdumb\b", r"\buseless\b", r"\bshut up\b"
    ]

    if any(re.search(pattern, query) for pattern in greeting_patterns):
        return "greeting"
    if any(re.search(pattern, query) for pattern in thank_you_patterns):
        return "thank_you"
    if any(re.search(pattern, query) for pattern in small_talk_patterns):
        return "small_talk"
    if any(re.search(pattern, query) for pattern in unrelated_patterns):
        return "unrelated"
    if any(re.search(pattern, query) for pattern in goodbye_patterns):
        return "goodbye"
    if any(re.search(pattern, query) for pattern in rude_patterns):
        return "rude"

    return "normal"

# Function to Search for Relevant Answers
def search_text(query, top_k=2):
    query_embedding = np.array(model.encode(query, convert_to_numpy=True)).astype("float32").reshape(1, -1)
    distances, indices = index.search(query_embedding, top_k)

    results = []
    for idx in indices[0]:
        if idx >= 0:
            results.append(metadata[idx])
    
    return results

# Serve HTML Page
@app.route("/")
def home():
    return render_template("index.html")

import re
@app.route("/query", methods=["POST"])

def query_pdf():
    query = request.json.get("query")

    query_type = categorize_query(query)

    if query_type == "greeting":
        return jsonify({"text": "Hello! How can I assist you with Exelsys EasyHR?", "images": []})

    if query_type == "thank_you":
        return jsonify({"text": "You're welcome! How can I assist you further?", "images": []})

    if query_type == "small_talk":
        return jsonify({"text": "I'm here to assist with Exelsys EasyHR. How can I help?", "images": []})

    if query_type == "unrelated":
        return jsonify({"text": "I'm here to assist with Exelsys easyHR queries only.", "images": []})

    if query_type == "vague":
        return jsonify({"text": "Could you please provide more details?", "images": []})

    if query_type == "goodbye":
        return jsonify({"text": "You're welcome! Have a great day!", "images": []})

    if query_type == "rude":
        return jsonify({"text": "I'm here to assist you professionally.", "images": []})

    # Search for relevant PDF content using FAISS
    results = search_text(query, top_k=3)

    if not results:
        return jsonify({"text": "No relevant results found in the PDF.", "images": []})

    # Merge multiple text results
    retrieved_text = "\n\n---\n\n".join([res["text"] for res in results])
    print(retrieved_text)
    
    prompt_extract = PromptTemplate.from_template(
        """
        ### YOU ARE AN EXELSYS EASYHR GUIDE ASSISTANT:
        ### INSTRUCTIONS:
        ### YOU ARE AN EXELSYS EASYHR GUIDE ASSISTANT:
        ### INSTRUCTIONS:
        - Your job is to provide step-by-step guidance for the following user query.
        - Check user query and provide relevant information from the Exelsys easyHR PDF.
        - Only provide image path if they are relevant to or you response and user query.
        - Only format the text response in html tags in steps form and not image paths. Don't provide headings in big tags. 
        - Do not add questions to response.
        - Base your response **only** on the retrieved context from the PDF.
        - If no relevant information is found, simply respond with: "Not found."
        - If the user greets you (e.g., "Hello", "Hi", "Good morning"), respond politely but keep it brief.
        - If the query is unrelated to Exelsys easyHR, respond with: "I'm here to assist with Exelsys easyHR queries only." 
        
        Each step may have an associated image that should be placed **immediately after** the step description.

        ### **Formatting Rules:**
        1. Each step should be numbered (`1., 2., 3., etc.`).
        2. If a step has an associated image, include the Markdown image syntax **right after** the step description.
        3. Do NOT list all images at the bottom; place them **inline** with their respective steps.
        4. If no image is available for a step, simply move to the next step.

        ### Example Format:
        1. Step 1 Text (description of the step).
        <img src"static/output_images/step1.png" class="inline-image" onclick="expandImage(this)">
        
        2. Step 2 Text (description of the step).
        
        3. Step 3 Text (description of the step).
        <img src"static/output_images/step2.png" class="inline-image" onclick="expandImage(this)"> 
        
        ### USER QUERY:
        {query}

        ### CONTEXT FROM PDF:
        {retrieved_text}

        ### ANSWER:
        """
    )

    # Chain the prompt with ChatGroq
    chain_extract = prompt_extract | llm
    chat_response = chain_extract.invoke({"query": query, "retrieved_text": retrieved_text})

    # Convert response to string
    response_text = str(chat_response.content)

    # Final response JSON
    response = {
        "text": response_text  # Store multiple images in a list
    }
    print(response)

    return jsonify(response)



if __name__ == "__main__":
    app.run(debug=True)