src/streamlit_app.py

# ==========================================================
# streamlit_app.py — Stable Layout + Multilingual Enhancement (Hindi + English)
# ==========================================================
import os
import re
import streamlit as st
import torch

# ==========================================================
# ✅ PAGE CONFIGS
# ==========================================================
st.set_page_config(page_title="Enterprise Knowledge Assistant", layout="wide")
print("CUDA available:", torch.cuda.is_available())

# ==========================================================
# ⚙️ CACHE SETUP
# ==========================================================
CACHE_DIR = "/tmp/hf_cache"
os.makedirs(CACHE_DIR, exist_ok=True)
os.environ.update({
    "HF_HOME": CACHE_DIR,
    "TRANSFORMERS_CACHE": CACHE_DIR,
    "HF_DATASETS_CACHE": CACHE_DIR,
    "HF_MODULES_CACHE": CACHE_DIR,
})

# ==========================================================
# 📦 IMPORTS
# ==========================================================
from ingestion import extract_text_from_pdf, chunk_text
from vectorstore import build_faiss_index
from qa import retrieve_chunks, generate_answer, cache_embeddings, embed_chunks, genai_generate

# ==========================================================
# 🧠 LANGUAGE DETECTION HELPER (Fast, No Dependencies)
# ==========================================================
from langdetect import detect

def detect_language(text_sample: str) -> str:
    """
    Detects Hindi (Devanagari) or English.
    Returns "hi" for Hindi and "en" for English.
    """
    try:
        # Quick Unicode-based detection for Hindi
        if re.search(r"[\u0900-\u097F]", text_sample):
            return "hi"

        # Fallback to langdetect
        lang = detect(text_sample)
        return "hi" if lang.startswith("hi") else "en"
    except Exception:
        return "en"



# ==========================================================
# 🧠 SMART SUGGESTION GENERATOR — bilingual (Hindi + English)
# ==========================================================
def generate_dynamic_suggestions_from_toc(toc, chunks, doc_name="Document", doc_lang="en"):
    """
    Generates 5-7 short, natural questions from TOC + a sample of chunks.
    If doc_lang == "hi", the prompt asks the model to return questions in Hindi.
    """
    if not toc or not chunks:
        # sensible bilingual fallback
        return ["How do I start using this guide?", "What does this document cover?"] if doc_lang != "hi" else [
            "मैं इस गाइड का उपयोग कैसे शुरू करूँ?",
            "यह दस्तावेज़ क्या कवर करता है?"
        ]

    # Build candidate titles from TOC
    titles = []
    for sec, raw_title in toc:
        title = re.sub(r"^\s*[\dA-Za-z.\-]+\s*", "", raw_title)
        title = re.sub(r"\.{2,}\s*\d+$", "", title).strip()
        if 4 < len(title) < 120:
            titles.append(title)

    context_sample = " ".join(chunks[:3])[:4000]

    # Choose language-aware prompt
    if str(doc_lang).startswith("hi"):
        prompt = f"""
आप एक सामग्री सहायक हैं। नीचे दिए गए तालिका-समाचार (Table of Contents) और दस्तावेज़ के नमूना पाठ के आधार पर 5 से 7 संक्षिप्त, साफ़ और मानवीय प्रश्न बनाइए।
प्रत्येक प्रश्न हिंदी में होना चाहिए, 18 शब्दों से कम, और प्रश्न चिह्न "?" के साथ समाप्त होना चाहिए। प्रश्न केवल दस्तावेज़ से प्रेरित हों — नई जानकारी इजाद न करें।

दस्तावेज़: "{doc_name}"

TABLE OF CONTENTS:
{chr(10).join(['- ' + t for t in titles[:8]])}

SAMPLE TEXT:
{context_sample}

आउटपुट: हर प्रश्न को नई लाइन पर लिखें, किसी भी क्रम चिन्ह के साथ (1., -, •) चलेगा। केवल प्रश्न लिखें।
"""
    else:
        prompt = f"""
You are a content assistant. Based on the Table of Contents and the sample document text below, generate 5–7 short, natural user-facing questions.
Each question should be in English, <18 words, and end with a question mark.
Document: "{doc_name}"

TABLE OF CONTENTS:
{chr(10).join(['- ' + t for t in titles[:8]])}

SAMPLE TEXT:
{context_sample}

Output: Put one question per line. Do not invent facts — base questions on the document.
"""

    try:
        ai_response = genai_generate(prompt)

        # Normalize response to lines
        lines = [ln.strip() for ln in ai_response.splitlines() if ln.strip()]

        # Heuristics to extract candidate questions
        candidates = []
        for ln in lines:
            # remove bullet/ordinal prefixes like "1.", "-", "•"
            ln_clean = re.sub(r"^[\-\u2022\*\d\.\)\s]+", "", ln).strip()

            # if line already ends with a question mark, keep it
            if ln_clean.endswith("?"):
                q = ln_clean
            else:
                # sometimes model returns without "?" but as a question — add "?" if short and starts with W/H or Hindi question words
                if (len(ln_clean.split()) < 18) and re.match(r"(?i)^(what|how|why|where|who|when|which)\b", ln_clean):
                    q = ln_clean + "?"
                # Hindi question words heuristic
                elif re.match(r"^(क्या|क्यों|कैसे|कहाँ|कौन|किस|कब)\b", ln_clean):
                    q = ln_clean if ln_clean.endswith("?") else ln_clean + "?"
                else:
                    # skip lines that don't look like questions
                    continue

            # length/filter
            q = q.strip()
            if 8 <= len(q) <= 140:
                candidates.append(q)

        # dedupe while preserving order
        seen = set()
        final = []
        for q in candidates:
            key = q.lower()
            if key not in seen:
                seen.add(key)
                final.append(q)

        # If we ended up with none, fallback to naive generation from titles
        if not final:
            # form simple question templates from titles
            for t in titles[:7]:
                if str(doc_lang).startswith("hi"):
                    cand = t.rstrip(".") + " के बारे में क्या जानना चाहिए?"
                else:
                    cand = "What should I know about " + t.rstrip(".") + "?"
                final.append(cand)
        # limit to 7
        return final[:7]

    except Exception as e:
        # graceful bilingual fallback
        if str(doc_lang).startswith("hi"):
            return [
                "इस दस्तावेज़ को कैसे शुरू करूँ?",
                "इस दस्तावेज़ का मुख्य उद्देश्य क्या है?",
                "प्रमुख हिस्से कौन से हैं?"
            ]
        else:
            return ["How do I start using this guide?", "What does this document cover?"]


# ==========================================================
# 🎨 STYLING — MINIMAL ENTERPRISE DESIGN
# ==========================================================
st.markdown("""
<style>
div.block-container {padding-top: 1.2rem; max-width: 1080px;}
h1, h2, h3 {color: #f3f4f6; font-weight: 600;}
.suggest-chip {
    background: #0f1724;
    border: 1px solid #374151;
    border-radius: 14px;
    color: #e6eef8;
    padding: 8px 12px;
    cursor: pointer;
    font-size: 13px;
    margin: 6px 6px 6px 0;
    display: inline-block;
    transition: background 0.2s, transform 0.1s;
}
.suggest-chip:hover {background: #1e3a8a; transform: translateY(-2px);}
.answer-box {
    background: linear-gradient(180deg,#0b1220,#071027);
    border-left: 4px solid #3b82f6;
    border-radius: 8px;
    padding: 16px 18px;
    color: #e6eef8;
    margin-top: 12px;
    box-shadow: 0 4px 14px rgba(0,0,0,0.35);
}
.stTextInput > div > div > input {
    background-color: #0f172a !important;
    color: #f1f5f9 !important;
    border-radius: 6px !important;
    border: 1px solid #334155 !important;
    padding: 8px 10px !important;
    font-size: 15px !important;
}
.stTextInput > label {font-weight: 500;}
.small-link {
    font-size: 13px;
    color: #60a5fa;
    cursor: pointer;
}
</style>
""", unsafe_allow_html=True)

# ==========================================================
# 🧭 SIDEBAR
# ==========================================================
with st.sidebar:
    st.markdown("### 🧭 Response Style")
    mode = st.radio(
        "",
        ("Strict (Document-only)", "Extended (Document + general)"),
        index=0,
        help="Strict = answers only from the uploaded document. Extended = may include related general info.",
    )

    st.markdown("---")
    show_dev = st.checkbox("Show advanced settings (for developers)", value=False)
    if show_dev:
        st.markdown("### ⚙️ Developer Options")
        chunk_size = st.slider("Chunk Size", 200, 1500, 1000, step=50)
        overlap = st.slider("Chunk Overlap", 50, 200, 120, step=10)
        top_k = st.slider("Top K Results", 1, 10, 7)
    else:
        chunk_size, overlap, top_k = 1000, 120, 5

    st.markdown("---")
    st.caption("✨ Built by Shubham Sharma")

# ==========================================================
# 🧠 SESSION STATE
# ==========================================================
for key, val in {
    "user_query_input": "",
    "show_more": False,
    "selected_suggestion": None,
    "query_suggestions_fixed": None,
    "last_doc": None,
    "doc_lang": "en",  # 🆕 store document language
}.items():
    if key not in st.session_state:
        st.session_state[key] = val

def set_user_query(q, idx):
    st.session_state["user_query_input"] = q
    st.session_state["selected_suggestion"] = idx
    st.experimental_rerun()

# ==========================================================
# 📄 MAIN SECTION
# ==========================================================
st.title("📄 Enterprise Knowledge Assistant")
st.caption("Query SAP documentation and enterprise PDFs — powered by reasoning and retrieval.")

doc_choice = st.radio("Select a document:", ["-- Select --", "Sample PDF", "Upload Custom PDF"], index=0)

# ==========================================================
# 📂 DOCUMENT HANDLING
# ==========================================================
if doc_choice == "-- Select --":
    st.info("⬅️ Select or upload a document to begin.")
else:
    if doc_choice == "Sample PDF":
        temp_path = os.path.join(os.path.dirname(__file__), "sample.pdf")
        st.success("📘 Sample document loaded successfully — you can start asking your questions below.")
    else:
        uploaded_file = st.file_uploader("", type="pdf", label_visibility="collapsed")
        if uploaded_file:
            temp_path = os.path.join("/tmp", uploaded_file.name)
            with open(temp_path, "wb") as f:
                f.write(uploaded_file.getbuffer())
            st.success("✅ Document processed successfully — you can start asking your questions below.")
        else:
            temp_path = None

    if temp_path:
        with st.spinner("🔍 Processing document..."):
            text, toc, toc_source = extract_text_from_pdf(temp_path)
            chunks = chunk_text(text, chunk_size=chunk_size, overlap=overlap)

            # 🌐 Detect document language (robust multilingual)
            doc_sample = " ".join(chunks[:3])[:3000]
            doc_lang = detect_language(doc_sample)
            st.session_state["doc_lang"] = doc_lang
            lang_label = "Hindi" if doc_lang.startswith("hi") else "English"
            st.caption(f"🈹 Detected document language: {lang_label}")

        with st.spinner("⚙️ Building search index..."):
            embeddings = cache_embeddings(os.path.basename(temp_path), chunks, embed_chunks)
            index = build_faiss_index(embeddings)

        doc_name = os.path.basename(temp_path)
        if st.session_state["last_doc"] != doc_name:
            query_suggestions = generate_dynamic_suggestions_from_toc(toc, chunks, doc_name, doc_lang)
            st.session_state["query_suggestions_fixed"] = query_suggestions
            st.session_state["last_doc"] = doc_name
            st.session_state["user_query_input"] = ""
            st.session_state["selected_suggestion"] = None
            st.session_state["show_more"] = False
            st.rerun()
        else:
            query_suggestions = st.session_state["query_suggestions_fixed"]

        # ----------------------------------------------------------
        # 💬 ASK SECTION
        # ----------------------------------------------------------
        st.markdown("### 💬 Ask the Assistant")

        if query_suggestions:
            visible = query_suggestions if st.session_state["show_more"] else query_suggestions[:3]
            cols = st.columns(min(3, len(visible)))
            for i, q in enumerate(visible):
                if cols[i % 3].button(f"💬 {q}", key=f"sugg_{i}"):
                    set_user_query(q, i)

            toggle_text = "Show less ▲" if st.session_state["show_more"] else "Show more ▼"
            if st.button(toggle_text, help="Show or hide more suggestions"):
                st.session_state["show_more"] = not st.session_state["show_more"]
                st.rerun()

        user_query = st.text_input("Type your question or click one above:", key="user_query_input")

        # ----------------------------------------------------------
        # 💡 RESPONSE SECTION
        # ----------------------------------------------------------
        if user_query.strip():
            reasoning_mode = mode == "Extended (Document + general)"
            with st.spinner("💭 Generating your answer..."):
                retrieved = retrieve_chunks(user_query, index, chunks, top_k=top_k, embeddings=embeddings)
                doc_lang = st.session_state.get("doc_lang", "en")
                print("🧠 Document language used for GPT prompt:", doc_lang)
                answer = generate_answer(user_query, retrieved, reasoning_mode=reasoning_mode, doc_lang=doc_lang)

            st.markdown("### 🤖 Assistant’s Answer")

            if not reasoning_mode and not answer.startswith("⚠️"):
                answer = re.sub(r"\*\*(.*?)\*\*", r"\1", answer)
                answer = re.sub(r"(^|\n)-\s*", r"\1<br>• ", answer)
            st.markdown(f"<div class='answer-box'>{answer}</div>", unsafe_allow_html=True)

            with st.expander("📘 Supporting Context"):
                for i, r in enumerate(retrieved, start=1):
                    st.markdown(f"**Chunk {i}:** {r}")

            if toc:
                with st.expander("📚 Explore Document Sections"):
                    toc_text = "\n".join([f"{sec}. {title}" for sec, title in toc])
                    st.text_area("", toc_text, height=140)

            with st.expander("📄 Document Preview"):
                st.text_area("", text[:1000], height=140)
                st.caption(f"{len(chunks)} chunks processed.")