app.py

import streamlit as st
import os
import tempfile
import asyncio
import nest_asyncio
import shutil
import torch
import gc
import networkx as nx
from pyvis.network import Network
import streamlit.components.v1 as components
from sentence_transformers import SentenceTransformer
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from lightrag import LightRAG, QueryParam
from lightrag.utils import EmbeddingFunc
from langchain_community.document_loaders import PyPDFLoader

# Allow nested loops for asyncio
nest_asyncio.apply()

# --- Page Config ---
st.set_page_config(page_title="LightRAG Case Study Guide", layout="wide")

st.title("📚 LightRAG Case Study Guide for Students")
st.markdown("""
This application implements the **LightRAG** framework using the `lightrag-hku` library.
It uses a dual-level retrieval system (Low-level & High-level) to maintain relational context.
""")

# --- Sidebar Parameters ---
with st.sidebar:
    st.header("RAG Parameters")
    chunk_size = st.slider("Chunk Token Size", 100, 1000, 512)
    chunk_overlap = st.slider("Chunk Overlap Token Size", 0, 200, 50)
    query_mode = st.selectbox("Query Mode", ["hybrid", "naive", "local", "global"], index=0)
    
    st.divider()
    st.info("Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0(4-bit Quantized)")
    st.success("Optimized for 16GB Memory Limit")

# --- Model Loading ---
@st.cache_resource
def load_models():
    # Clear memory before loading
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    st.info("🧠 Loading Embedding Model (CPU)...")
    embedding_model = SentenceTransformer("all-MiniLM-L6-v2", device="cpu")
    
    st.info("⏳ Loading LLM (Qwen2.5-3B-Instruct)...")
    # Using Qwen2.5-3B-Instruct as a balance between performance and memory
    model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
    
    device = "cuda" if torch.cuda.is_available() else "cpu"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    
    # Always use 4-bit quantization to stay within 16GB, even on CPU if possible
    # Note: bitsandbytes 4-bit is primarily for CUDA. For CPU, we'll use float32 with low_cpu_mem_usage.
    if device == "cuda":
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True, 
            bnb_4bit_compute_dtype=torch.float16,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_use_double_quant=True
        )
        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            quantization_config=bnb_config,
            device_map="auto",
        )
    else:
        # On CPU, we use float32 but ensure low_cpu_mem_usage. 3B model in float32 is ~6GB.
        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            torch_dtype=torch.float32,
            low_cpu_mem_usage=True,
        ).to("cpu")
        
    return embedding_model, tokenizer, model

embedding_model, tokenizer, model = load_models()

# --- LightRAG Functions ---
async def embed_func(texts):
    return embedding_model.encode(texts, convert_to_numpy=True)

def sync_llm_call(prompt, kwargs):
    system_prompt = kwargs.get("system_prompt") or "You are a helpful assistant."
    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": prompt},
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    
    if inputs["input_ids"].shape[1] > 1500:
        inputs["input_ids"] = inputs["input_ids"][:, :1500]
        inputs["attention_mask"] = inputs["attention_mask"][:, :1500]
    
    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=512,
            do_sample=False,
            temperature=None,
            top_p=None,
        )
    
    new_tokens = output_ids[0][inputs["input_ids"].shape[1]:]
    return tokenizer.decode(new_tokens, skip_special_tokens=True).strip()

async def llm_func(prompt, **kwargs):
    safe_kwargs = {k: v for k, v in kwargs.items() if isinstance(v, (str, int, float, bool, type(None)))}
    loop = asyncio.get_event_loop()
    return await loop.run_in_executor(None, sync_llm_call, prompt, safe_kwargs)

# --- App Logic ---
WORKING_DIR = "./lightrag_storage"
if not os.path.exists(WORKING_DIR):
    os.makedirs(WORKING_DIR)

# Initialize RAG instance
rag = LightRAG(
    working_dir=WORKING_DIR,
    embedding_func=EmbeddingFunc(
        embedding_dim=384,
        max_token_size=256,
        func=embed_func,
    ),
    llm_model_func=llm_func,
    chunk_token_size=chunk_size,
    chunk_overlap_token_size=chunk_overlap,
)

uploaded_file = st.file_uploader("Upload a PDF document", type="pdf")

if uploaded_file:
    # Use a persistent temp file path to avoid unlinking issues across button clicks
    tmp_path = os.path.join(tempfile.gettempdir(), "uploaded_doc.pdf")
    with open(tmp_path, "wb") as f:
        f.write(uploaded_file.getvalue())

    if st.button("Index Document"):
        with st.status("Processing...", expanded=True) as status:
            st.write("📄 Loading PDF...")
            loader = PyPDFLoader(tmp_path)
            pages = loader.load()
            documents = [p.page_content for p in pages]
            
            st.write(f"✂️ Indexing {len(documents)} pages into LightRAG...")
            
            async def run_indexing():
                await rag.initialize_storages()
                await rag.ainsert(documents)
            
                # Clear memory after indexing
                gc.collect()
                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
            
            asyncio.run(run_indexing())
            st.write("✅ Indexing complete!")
            status.update(label="Ready for Questions!", state="complete", expanded=False)
            st.session_state['indexed'] = True

    if st.session_state.get('indexed'):
        tab1, tab2 = st.tabs(["💬 Q&A Interface", "🕸️ Knowledge Graph"])

        with tab1:
            query = st.text_input("Ask a question:")
            if query:
                with st.spinner("Generating answer..."):
                    async def run_query():
                        return await rag.aquery(query, param=QueryParam(mode=query_mode))
                    
                    answer = asyncio.run(run_query())
                    st.markdown("### 🤖 Answer")
                    st.write(answer)

        with tab2:
            st.markdown("### Knowledge Graph")
            graph_path = os.path.join(WORKING_DIR, "graph_chunk_entity_relation.graphml")
            if os.path.exists(graph_path):
                try:
                    G = nx.read_graphml(graph_path)
                    if len(G.nodes) > 100:
                        st.warning("Graph is large. Showing a subgraph of the first 100 nodes.")
                        nodes = list(G.nodes)[:100]
                        G = G.subgraph(nodes)
                    
                    net = Network(height="500px", width="100%", bgcolor="#ffffff", font_color="black")
                    net.from_nx(G)
                    with tempfile.NamedTemporaryFile(delete=False, suffix=".html") as tmp_html:
                        net.save_graph(tmp_html.name)
                        with open(tmp_html.name, 'r', encoding='utf-8') as f:
                            components.html(f.read(), height=550)
                        os.unlink(tmp_html.name)
                except Exception as e:
                    st.error(f"Error loading graph: {e}")
            else:
                st.info("Knowledge graph file not found. It will be generated during indexing.")
else:
    st.info("Please upload a PDF to start.")
    if 'indexed' in st.session_state:
        st.session_state['indexed'] = False