app.py

import os
import docx
import pandas as pd
import numpy as np
import streamlit as st
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer
import faiss
from groq import Groq

# ==========================================================
# GROQ API KEY (use HF Secrets)
# ==========================================================
os.environ["GROQ_API_KEY"] = os.getenv("API")

# ==========================================================
# STREAMLIT UI
# ==========================================================
st.set_page_config(page_title="Word RAG App", layout="wide")
st.title("📄 Word Document RAG")

uploaded_file = st.file_uploader(
    "Upload a Word document (.docx)",
    type=["docx"]
)

# ==========================================================
# WORD TEXT EXTRACTION (UNCHANGED)
# ==========================================================
def read_word(doc_path):
    doc = docx.Document(doc_path)
    text = "\n\n".join([p.text for p in doc.paragraphs if p.text.strip() != ""])
    return [{"page": 1, "text": text}]

# ==========================================================
# CORE RAG FUNCTIONS (UNCHANGED)
# ==========================================================
def chunk_text(pages, chunk_size=800):
    chunks = []
    for page in pages:
        paragraphs = page["text"].split("\n\n")
        buffer = ""
        for para in paragraphs:
            if len(buffer) + len(para) <= chunk_size:
                buffer += " " + para
            else:
                chunks.append({"page": page["page"], "text": buffer.strip()})
                buffer = para
        if buffer:
            chunks.append({"page": page["page"], "text": buffer.strip()})
    return chunks

def tokenize_chunks(chunks, model_name="sentence-transformers/all-mpnet-base-v2"):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    return [tokenizer(c["text"], truncation=True)["input_ids"] for c in chunks]

def create_embeddings(chunks, model_name="allenai/specter"):
    embedder = SentenceTransformer(model_name)
    texts = [c["text"] for c in chunks]
    embeddings = embedder.encode(texts, show_progress_bar=False)
    return embedder, np.array(embeddings)

def store_embeddings(embeddings):
    faiss.normalize_L2(embeddings)
    dim = embeddings.shape[1]
    index = faiss.IndexFlatIP(dim)
    index.add(embeddings)
    return index

def retrieve_chunks(query, embedder, index, chunks, top_k=None):
    if not top_k:
        top_k = min(20, len(chunks))
    query_vec = embedder.encode([query])
    faiss.normalize_L2(query_vec)
    scores, indices = index.search(query_vec, top_k)
    return [chunks[i] for i in indices[0]]

def build_safe_context(retrieved_chunks, max_chars=12000):
    context = ""
    used = 0
    for c in retrieved_chunks[:3]:
        block = f"(Page {c['page']}) {c['text']}\n\n"
        context += block
        used += len(block)
    for c in retrieved_chunks[3:]:
        block = f"(Page {c['page']}) {c['text']}\n\n"
        if used + len(block) > max_chars:
            break
        context += block
        used += len(block)
    return context

def generate_answer(query, context):
    client = Groq()
    prompt = f"""
    You are a document-based assistant.
    Use the context to answer the question clearly.
    If the answer is partially available, summarize it.
    If the answer is not present, say 'Not found in the document'.

    Context:
    {context}

    Question:
    {query}
    """
    response = client.chat.completions.create(
        model="llama-3.1-8b-instant",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.3
    )
    return response.choices[0].message.content

# ==========================================================
# APP LOGIC
# ==========================================================
if uploaded_file:
    with st.spinner("📄 Reading document..."):
        temp_path = "/tmp/uploaded.docx"
        with open(temp_path, "wb") as f:
            f.write(uploaded_file.getbuffer())

        pages = read_word(temp_path)

    with st.spinner("✂️ Chunking & embedding document..."):
        chunks = chunk_text(pages)
        tokenize_chunks(chunks)
        embedder, embeddings = create_embeddings(chunks)
        index = store_embeddings(embeddings)

    st.success("✅ Document indexed successfully")

    query = st.text_input("❓ Ask a question")

    if query:
        with st.spinner("🤖 Generating answer..."):
            retrieved_chunks = retrieve_chunks(query, embedder, index, chunks)
            context = build_safe_context(retrieved_chunks)
            answer = generate_answer(query, context)

        st.markdown("### ✅ Answer")
        st.write(answer)