src/streamlit_app.py

import streamlit as st
from llama_index.core import VectorStoreIndex, Document, Settings, SimpleDirectoryReader, StorageContext
from llama_index.core.node_parser import SemanticSplitterNodeParser
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
from llama_index.llms.ollama import Ollama
from llama_index.retrievers.bm25 import BM25Retriever
from llama_index.core.retrievers import QueryFusionRetriever
from llama_index.core.retrievers.fusion_retriever import FUSION_MODES
from llama_index.core.chat_engine import CondensePlusContextChatEngine
from llama_index.core.memory import ChatMemoryBuffer
import httpx
import os
import datetime
from huggingface_hub import HfApi, CommitScheduler
from pathlib import Path
import json
import uuid

# --- 1. CONFIGURATION ---
RESEARCHER_NAME = "Enoch Hyunwook Kang"
OLLAMA_BASE_URL = "https://researchbot.share.zrok.io" 
OLLAMA_MODEL = "qwen3:8b" 

# --- 2. LOGGING SETUP (Hugging Face Dataset) ---
# Create a private dataset on HF (e.g., "ehwkang/researchbot-logs") first!
LOG_DATASET = "ehwkang/researchbot-logs" 
LOG_FILE = "qna_logs.jsonl"
scheduler = CommitScheduler(
    repo_id=LOG_DATASET,
    repo_type="dataset",
    folder_path="logs",
    path_in_repo="data",
    every=10 # Upload every 10 minutes (or on shutdown)
)

def log_interaction(question, answer):
    # Determine the log file path
    log_path = Path("logs") / LOG_FILE
    log_path.parent.mkdir(parents=True, exist_ok=True)
    
    entry = {
        "timestamp": datetime.datetime.now().isoformat(),
        "session_id": st.session_state.get("session_id"),
        "question": question,
        "answer": str(answer)
    }
    
    with scheduler.lock:
        with log_path.open("a") as f:
            f.write(json.dumps(entry) + "\n")

# --- 3. CUSTOM OLLAMA CLIENT ---
class CustomOllama(Ollama):
    def _get_client(self):
        return httpx.Client(
            base_url=self.base_url, 
            timeout=120.0, 
            headers={"skip_zrok_interstitial": "true"}
        )

# --- 4. SETUP ---
st.set_page_config(page_title=f"{RESEARCHER_NAME}'s Research", layout="centered")

if "session_id" not in st.session_state:
    st.session_state.session_id = str(uuid.uuid4())

# Initialize Models
try:
    # Embedding Model (Runs on HF CPU - lightweight)
    embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")
    Settings.embed_model = embed_model
    
    # LLM (Runs on your Local GPU)
    Settings.llm = CustomOllama(
        model=OLLAMA_MODEL,
        base_url=OLLAMA_BASE_URL,
        request_timeout=120.0,
        context_window=8192,  # 8k is usually enough for RAG
        temperature=0.3       # Lower temp for factual research answers
    )
except Exception as e:
    st.error(f"Configuration Error: {e}")

# --- 5. INTELLIGENT INDEXING (Semantic + Hybrid) ---
@st.cache_resource
def load_resources():
    script_dir = os.path.dirname(os.path.abspath(__file__))
    
    # A. Load CV
    cv_text = ""
    cv_path = os.path.join(script_dir, "CV.txt")
    if os.path.exists(cv_path):
        with open(cv_path, "r", encoding="utf-8") as f:
            cv_text = f.read()

    # B. Load Papers & Build Index
    data_dir = os.path.join(script_dir, "data")
    if not os.path.exists(data_dir):
        return cv_text, None

    documents = SimpleDirectoryReader(data_dir, required_exts=[".txt"], recursive=True).load_data()
    
    if not documents:
        return cv_text, None

    # SOTA 1: Semantic Chunking (Splits by meaning, not just line count)
    # Note: This runs on CPU (HF Spaces), so it might take 30-60s on boot.
    splitter = SemanticSplitterNodeParser(
        buffer_size=1, 
        breakpoint_percentile_threshold=95, 
        embed_model=embed_model
    )
    
    nodes = splitter.get_nodes_from_documents(documents)
    
    # Create Vector Index
    vector_index = VectorStoreIndex(nodes)
    
    return cv_text, vector_index, nodes

cv_content, vector_index, all_nodes = load_resources()

# --- 6. HYBRID RETRIEVER & CHAT ENGINE ---
def get_chat_engine():
    if not vector_index:
        return None

    # SOTA 2: Hybrid Retrieval (Vector + BM25)
    # 1. Vector Search (Semantic understanding)
    vector_retriever = vector_index.as_retriever(similarity_top_k=5)
    
    # 2. BM25 Search (Keyword precision - crucial for specific algorithm names)
    bm25_retriever = BM25Retriever.from_defaults(nodes=all_nodes, similarity_top_k=5)
    
    # 3. Fusion (Combine results)
    retriever = QueryFusionRetriever(
        [vector_retriever, bm25_retriever],
        similarity_top_k=5,
        num_queries=1,
        mode=FUSION_MODES.RECIPROCAL_RANK,  # <--- USE ENUM (safest) 
        # OR use mode="reciprocal_rerank" (note the extra 're')
        use_async=False,
    )

    # SOTA 3: CondensePlusContext
    # Handles: "What is its accuracy?" -> "What is the accuracy of [Previous Topic]?"
    memory = ChatMemoryBuffer.from_defaults(token_limit=4000)
    
    system_prompt = (
        f"You are {RESEARCHER_NAME}. Answer questions about your research based ONLY on the provided context. "
        f"If the answer is not in the context, say you don't know. "
        f"Here is your CV for biographical context:\n{cv_content}"
    )

    return CondensePlusContextChatEngine.from_defaults(
        retriever=retriever,
        llm=Settings.llm,
        memory=memory,
        system_prompt=system_prompt,
        verbose=True
    )

chat_engine = get_chat_engine()

# --- 7. CHAT UI ---
if "messages" not in st.session_state:
    st.session_state.messages = [{"role": "assistant", "content": "Hello! Ask me about my research."}]

for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.write(message["content"])

if prompt := st.chat_input("Ask a question..."):
    st.session_state.messages.append({"role": "user", "content": prompt})
    with st.chat_message("user"):
        st.write(prompt)

    if chat_engine:
        with st.chat_message("assistant"):
            with st.spinner("Thinking..."):
                response = chat_engine.chat(prompt)
                st.write(str(response))
                st.session_state.messages.append({"role": "assistant", "content": str(response)})
                
                # Log to HF Dataset
                log_interaction(prompt, response)
    else:
        st.error("Index not loaded. Check 'data' folder.")