app.py

import streamlit as st
import streamlit.components.v1 as components
import os
import html
import re
import torch
from datasets import load_dataset
from langchain_ollama import OllamaLLM
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.chains import RetrievalQA
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.docstore.document import Document
from langchain_openai import ChatOpenAI
from langchain_together import Together

from langchain.prompts.chat import ChatPromptTemplate, SystemMessagePromptTemplate, HumanMessagePromptTemplate

# 🔹 Load and inject custom CSS
def load_css(file_name):
    with open(file_name) as f:
        st.markdown(f'<style>{f.read()}</style>', unsafe_allow_html=True)

load_css("styles.css")

# 🔹 System prompt
system_prompt = """
You are Adrega AI, a helpful assistant trained to answer questions based on internal company documentation.

Instructions:
- Be concise and professional and keep response length under 512 tokens if possible.
- Use markdown formatting when helpful.
- Never show the thought process or reasoning steps.
- If the user asks about a specific operation in the product, explain step by step how to do it.
- If the user greets you (e.g. “Hi”, “Hello”), respond with a friendly greeting and offer help. Do not reference documentation unless a question is asked.
- Reference the Help page from the hamburger menu when needed.
- The Field Reference module is not a physical module, but a description of fields in Adrega. Prioritize it when describing fields.
- Do not make assumptions. Use the manual as reference.
"""

# 🔹 Chat prompt template
chat_prompt = ChatPromptTemplate.from_messages([
    SystemMessagePromptTemplate.from_template(system_prompt),
    HumanMessagePromptTemplate.from_template("""
Context:
{context}

Question:
{question}
""")
])

# 🔹 Provider selection
#provider_options = ["Together", "HuggingFace", "Adrega"]
provider_options = ["Adrega", "Together"]
selected_provider = st.selectbox("Provider", provider_options)

# 🔹 Load and process dataset once
if "initialized" not in st.session_state:
    dataset = load_dataset("andreska/Adrega61Docs", split="train")

    def read_dataset(dataset):
        return "\n\n".join([
            f"Title: {item['title']}\nModule: {item['module']}\nContent: {item['content']}"
            for item in dataset
        ])

    raw_text = read_dataset(dataset)
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
    docs = text_splitter.split_documents([Document(page_content=raw_text)])

    # Smart device detection for HuggingFace Spaces
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    embedding_model = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-MiniLM-L6-v2",
        model_kwargs={'device': device}
    )
    db = FAISS.from_documents(docs, embedding_model)

    st.session_state.db = db
    st.session_state.docs = docs
    st.session_state.initialized = True

# 🔹 Provider-specific LLM setup
def get_llm(provider):
    # Validate provider input against allowed options
    allowed_providers = ["Adrega", "Together", "HuggingFace"]
    if provider not in allowed_providers:
        provider = "Adrega"  # Default fallback
    
    if provider == "Together":
        return Together(
            # Alternative models available for testing
            #model="meta-llama/Llama-3.2-11B-Vision-Free",
            model="deepseek-ai/DeepSeek-R1-Distilled",
            #model="flux-ai/FLUX.1-schnell-Free",
            max_tokens=2048
        )
    elif provider == "HuggingFace":
        return ChatOpenAI(
            base_url="https://router.huggingface.co/v1",
            api_key=os.environ["HF_API_KEY"],
            model="HuggingFaceTB/SmolLM3-3B:hf-inference",
        )
    else:  # Adrega
        return OllamaLLM(
            base_url=os.environ["ADREGA_URL"],
            #model="phi3:mini",  # Alternative lightweight model
            model="llama3:latest",
            streaming=False,
        )

# 🔹 Build the LLM and retriever
llm = get_llm(selected_provider)
retriever = st.session_state.db.as_retriever(search_kwargs={"k": 5})

# 🔹 UI Styling
st.markdown("""<style>
.scrollable-div { 
    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif !important; 
    max-height: 400px; 
    overflow-y: auto; 
    padding: 10px; 
    border: 1px solid #ccc; 
    background-color: #f9f9f9; 
}
input[data-testid="stTextInput"] { cursor: pointer; }
input[data-testid="stTextInput"]:focus { cursor: text; }
</style>
<script>
document.addEventListener('DOMContentLoaded', function() {
    const observer = new MutationObserver(function(mutations) {
        const input = document.querySelector('input[data-testid="stTextInput"]');
        if (input && !input.hasAttribute('data-select-all')) {
            input.setAttribute('data-select-all', 'true');
            input.addEventListener('click', function() {
                this.select();
            });
        }
    });
    observer.observe(document.body, { childList: true, subtree: true });
});
</script>""", unsafe_allow_html=True)



# 🔹 Initialize chat history
if "chat_history" not in st.session_state:
    st.session_state.chat_history = [{"role": "assistant", "content": "Hi! I am your Adrega AI assistant. How can I help you today?"}]

# 🔹 Build limited context from history
def build_context(history, max_chars=10000):
    context = []
    total_chars = 0
    for msg in reversed(history):
        if msg["role"] == "user":
            entry = f"You: {msg['content']}\n"
        else:
            entry = f"Adrega AI: {msg['content']}\n"
        total_chars += len(entry)
        if total_chars > max_chars:
            break
        context.insert(0, entry)
    return "\n".join(context)

def render_chat():
    
    html_content = """
    <style>
        .scrollable-div {
            font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif !important; 
            max-height: 400px;
            overflow-y: auto;
            padding: 10px;
            border: 1px solid #ccc;
            background-color: #f9f9f9;
        }
    </style>
    <div class="scrollable-div" id="chat-box">
    """

    for msg in st.session_state.chat_history:
        content = msg["content"]
        
        # Decode HTML entities first
        content = html.unescape(content)
        
        # Convert markdown to HTML
        content = content.replace('\n', '<br>')  # Line breaks
        content = content.replace('**', '<b>', 1).replace('**', '</b>', 1)
        while '**' in content:
            content = content.replace('**', '<b>', 1).replace('**', '</b>', 1)
        
        # Convert lists to HTML
        lines = content.split('<br>')
        processed_lines = []
        in_ul_list = False
        in_ol_list = False
        
        for line in lines:
            stripped = line.strip()
            # Check for numbered lists (1. 2. etc.)
            if re.match(r'^\d+\.\s+', stripped):
                if in_ul_list:
                    processed_lines.append('</ul>')
                    in_ul_list = False
                if not in_ol_list:
                    processed_lines.append('<ol>')
                    in_ol_list = True
                item_text = re.sub(r'^\d+\.\s+', '', stripped)
                processed_lines.append(f'<li>{item_text}</li>')
            # Check for bullet lists
            elif stripped.startswith('- ') or stripped.startswith('* '):
                if in_ol_list:
                    processed_lines.append('</ol>')
                    in_ol_list = False
                if not in_ul_list:
                    processed_lines.append('<ul>')
                    in_ul_list = True
                processed_lines.append(f'<li>{stripped[2:]}</li>')
            else:
                if in_ul_list:
                    processed_lines.append('</ul>')
                    in_ul_list = False
                if in_ol_list:
                    processed_lines.append('</ol>')
                    in_ol_list = False
                processed_lines.append(line)
        
        if in_ul_list:
            processed_lines.append('</ul>')
        if in_ol_list:
            processed_lines.append('</ol>')
        
        content = ''.join(processed_lines)
        
        if msg["role"] == "user":
            html_content += f'<div style="background-color:#e0e0e0; padding:8px; border-radius:5px; margin-bottom:6px;"><b>You:</b> {content}</div>'
        else:
            html_content += f'<div style="padding:8px; margin-bottom:6px;"><b>Adrega AI:</b> {content}</div>'

    html_content += """
    </div>
    <script>
        const chatBox = document.getElementById("chat-box");
        if (chatBox) {
            chatBox.scrollTop = chatBox.scrollHeight;
        }
    </script>
    """

    components.html(html_content, height=420, scrolling=False)

# 🔹 Handle user input
def handle_submit():
    user_input = st.session_state.user_input
    if user_input:
        context = build_context(st.session_state.chat_history)
        
        # Check if user needs documentation (questions or specific requests)
        needs_docs = any(word in user_input.lower() for word in ['how', 'what', 'where', 'when', 'why', '?']) or len(user_input.split()) > 5
        
        if needs_docs:
            # Retrieve relevant docs for complex queries
            docs = retriever.get_relevant_documents(user_input)
            doc_context = "\n\n".join([doc.page_content for doc in docs[:3]])
            if selected_provider == "Together":
                full_prompt = f"{system_prompt}\n\nConversation History:\n{context}\n\nDocumentation Context:\n{doc_context}\n\nUser: {user_input}\nAdrega AI:"
            else:
                full_prompt = f"{system_prompt}\n\nConversation History:\n{context}\n\nDocumentation Context:\n{doc_context}\n\nUser: {user_input}\n\nRespond directly as Adrega AI without showing your thought process:"
        else:
            # Simple conversational response without docs
            if selected_provider == "Together":
                full_prompt = f"{system_prompt}\n\nConversation History:\n{context}\n\nUser: {user_input}\nAdrega AI:"
            else:
                full_prompt = f"{system_prompt}\n\nConversation History:\n{context}\n\nUser: {user_input}\n\nAnswer directly without any reasoning or explanation of your thought process:"
        
        answer = llm.invoke(full_prompt)
        
        # Extract content from response object
        if hasattr(answer, 'content'):
            answer = answer.content
        elif isinstance(answer, dict) and 'content' in answer:
            answer = answer['content']
        
        # Clean up response for non-Together providers
        if selected_provider != "Together" and isinstance(answer, str):
            # Remove instruction-following sentences at the beginning
            sentences = answer.split('. ')
            clean_sentences = []
            
            for sentence in sentences:
                # Skip sentences that are clearly instruction-following
                if not any(phrase in sentence.lower() for phrase in [
                    'keep the response', 'as per the instructions', 'per the instructions',
                    'following the instructions', 'according to', 'make sure to',
                    'remember to', 'be sure to', 'don\'t forget to'
                ]):
                    clean_sentences.append(sentence.strip())
            
            if clean_sentences:
                answer = '. '.join(clean_sentences)
                # Clean up any remaining artifacts
                if answer.endswith('.'):
                    answer = answer[:-1] + '.'
            
            # Additional cleanup for remaining reasoning patterns
            lines = answer.split('\n')
            clean_lines = []
            for line in lines:
                if not any(phrase in line.lower() for phrase in ['okay,', 'let me', 'looking at', 'i need', 'i should', 'wait,', 'so the response', 'therefore']):
                    clean_lines.append(line)
            answer = '\n'.join(clean_lines).strip()

        # Update history
        st.session_state.chat_history.append({"role": "user", "content": user_input})
        st.session_state.chat_history.append({"role": "assistant", "content": answer})

# 🔹 Create container for chat display
chat_container = st.container()

# 🔹 Input field
st.text_input('Ask me a question', key='user_input', on_change=handle_submit)
if st.button("Ask"):
    handle_submit()    

# 🔹 Display chat history in container
with chat_container:
    render_chat()