streamlit_app.py

"""Streamlit chat interface for RAG application."""
import streamlit as st
import sys
import os
from datetime import datetime
import json
import pandas as pd
from typing import Optional
import warnings

# Suppress warnings
warnings.filterwarnings('ignore')
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

# Add parent directory to path
sys.path.append(os.path.dirname(os.path.abspath(__file__)))

# Check if running on HuggingFace Spaces
IS_HUGGINGFACE_SPACE = os.environ.get("SPACE_ID") is not None

from config import settings
from dataset_loader import RAGBenchLoader
from vector_store import ChromaDBManager, create_vector_store
try:
    from vector_store import QdrantManager, QDRANT_AVAILABLE
except ImportError:
    QDRANT_AVAILABLE = False
from llm_client import GroqLLMClient, OllamaLLMClient, RAGPipeline, create_llm_client
from trace_evaluator import TRACEEvaluator
from embedding_models import EmbeddingFactory
from chunking_strategies import ChunkingFactory
import uuid

# Background evaluation support
try:
    from background_evaluator import (
        create_job, start_background_job, load_job_status, 
        list_jobs, delete_job, STATUS_PENDING, STATUS_RUNNING, 
        STATUS_COMPLETED, STATUS_FAILED
    )
    BACKGROUND_EVAL_AVAILABLE = True
except ImportError:
    BACKGROUND_EVAL_AVAILABLE = False


# Page configuration
st.set_page_config(
    page_title="RAG Capstone Project",
    page_icon="🤖",
    layout="wide"
)

# Initialize session state
if "chat_history" not in st.session_state:
    st.session_state.chat_history = []

if "rag_pipeline" not in st.session_state:
    st.session_state.rag_pipeline = None

if "vector_store" not in st.session_state:
    st.session_state.vector_store = None

if "collection_loaded" not in st.session_state:
    st.session_state.collection_loaded = False

if "evaluation_results" not in st.session_state:
    st.session_state.evaluation_results = None

if "dataset_size" not in st.session_state:
    st.session_state.dataset_size = 10000

if "current_dataset" not in st.session_state:
    st.session_state.current_dataset = None

if "current_llm" not in st.session_state:
    st.session_state.current_llm = settings.llm_models[1]

if "selected_collection" not in st.session_state:
    st.session_state.selected_collection = None

if "available_collections" not in st.session_state:
    st.session_state.available_collections = []

if "dataset_name" not in st.session_state:
    st.session_state.dataset_name = None

if "collection_name" not in st.session_state:
    st.session_state.collection_name = None

if "embedding_model" not in st.session_state:
    st.session_state.embedding_model = None

if "groq_api_key" not in st.session_state:
    st.session_state.groq_api_key = ""

if "llm_provider" not in st.session_state:
    st.session_state.llm_provider = settings.llm_provider

if "ollama_model" not in st.session_state:
    st.session_state.ollama_model = settings.ollama_model

if "vector_store_provider" not in st.session_state:
    st.session_state.vector_store_provider = settings.vector_store_provider

if "qdrant_url" not in st.session_state:
    st.session_state.qdrant_url = settings.qdrant_url

if "qdrant_api_key" not in st.session_state:
    st.session_state.qdrant_api_key = settings.qdrant_api_key


def get_available_collections(provider: str = None):
    """Get list of available collections from vector store."""
    provider = provider or st.session_state.get("vector_store_provider", "chroma")
    
    try:
        if provider == "qdrant" and QDRANT_AVAILABLE:
            qdrant_url = st.session_state.get("qdrant_url") or settings.qdrant_url
            qdrant_api_key = st.session_state.get("qdrant_api_key") or settings.qdrant_api_key
            if qdrant_url and qdrant_api_key:
                vector_store = create_vector_store("qdrant", url=qdrant_url, api_key=qdrant_api_key)
                collections = vector_store.list_collections()
                return collections
            return []
        else:
            vector_store = ChromaDBManager(settings.chroma_persist_directory)
            collections = vector_store.list_collections()
            return collections
    except Exception as e:
        print(f"Error getting collections: {e}")
        return []


def main():
    """Main Streamlit application."""
    st.title("🤖 RAG Capstone Project")
    st.markdown("### Retrieval-Augmented Generation with TRACE Evaluation")
    
    # Show HuggingFace Spaces notice
    if IS_HUGGINGFACE_SPACE:
        st.info("🤗 Running on Hugging Face Spaces - Using Groq API (cloud-based LLM)")
    
    # Get available collections at startup
    available_collections = get_available_collections()
    st.session_state.available_collections = available_collections
    
    # Sidebar for configuration
    with st.sidebar:
        st.header("Configuration")
        
        # LLM Provider Selection - Disable Ollama on HuggingFace Spaces
        st.subheader("🔌 LLM Provider")
        
        if IS_HUGGINGFACE_SPACE:
            # Force Groq on HuggingFace Spaces (Ollama not available)
            st.caption("☁️ **Groq API** (Ollama unavailable on Spaces)")
            llm_provider = "groq"
            st.session_state.llm_provider = "groq"
        else:
            llm_provider = st.radio(
                "Choose LLM Provider:",
                options=["groq", "ollama"],
                index=0 if st.session_state.llm_provider == "groq" else 1,
                format_func=lambda x: "☁️ Groq API (Cloud)" if x == "groq" else "🖥️ Ollama (Local)",
                help="Groq: Cloud API with rate limits. Ollama: Local unlimited inference.",
                key="llm_provider_radio"
            )
            st.session_state.llm_provider = llm_provider
        
        # Provider-specific settings
        if llm_provider == "groq":
            st.caption("⚠️ Free tier: 30 requests/min")
            
            # On HuggingFace Spaces, check for API key in secrets first
            default_api_key = os.environ.get("GROQ_API_KEY", "") or settings.groq_api_key or ""
            
            # API Key input
            groq_api_key = st.text_input(
                "Groq API Key",
                type="password",
                value=default_api_key,
                help="Enter your Groq API key (or set GROQ_API_KEY in Spaces secrets)"
            )
            
            if IS_HUGGINGFACE_SPACE and not groq_api_key:
                st.warning("💡 Tip: Add GROQ_API_KEY to your Space secrets for persistence")
        else:
            # Ollama settings (only available locally)
            st.caption("✅ No rate limits - unlimited usage!")
            ollama_host = st.text_input(
                "Ollama Host",
                value=settings.ollama_host,
                help="Ollama server URL (default: http://localhost:11434)"
            )
            
            ollama_model = st.selectbox(
                "Select Ollama Model:",
                options=settings.ollama_models,
                index=settings.ollama_models.index(st.session_state.ollama_model) if st.session_state.ollama_model in settings.ollama_models else 0,
                key="ollama_model_selector"
            )
            st.session_state.ollama_model = ollama_model
            
            # Connection check button
            if st.button("🔍 Check Ollama Connection"):
                try:
                    import requests
                    response = requests.get(f"{ollama_host}/api/tags", timeout=5)
                    if response.status_code == 200:
                        models = response.json().get("models", [])
                        model_names = [m["name"] for m in models]
                        st.success(f"✅ Connected! Available models: {', '.join(model_names)}")
                    else:
                        st.error(f"❌ Connection failed: {response.status_code}")
                except Exception as e:
                    st.error(f"❌ Cannot connect to Ollama: {e}")
                    st.info("Make sure Ollama is running: `ollama serve`")
            
            groq_api_key = ""  # Not needed for Ollama
        
        st.divider()
        
        # Vector Store Provider Selection
        st.subheader("💾 Vector Store")
        
        if IS_HUGGINGFACE_SPACE:
            st.caption("☁️ Use **Qdrant Cloud** for persistent storage")
            vector_store_options = ["qdrant", "chroma"]
            default_idx = 0
        else:
            vector_store_options = ["chroma", "qdrant"]
            default_idx = 0
        
        vector_store_provider = st.radio(
            "Choose Vector Store:",
            options=vector_store_options,
            index=default_idx,
            format_func=lambda x: "☁️ Qdrant Cloud (Persistent)" if x == "qdrant" else "💾 ChromaDB (Local)",
            help="Qdrant: Cloud storage (persistent). ChromaDB: Local storage (ephemeral on Spaces).",
            key="vector_store_radio"
        )
        st.session_state.vector_store_provider = vector_store_provider
        
        # Qdrant settings
        if vector_store_provider == "qdrant":
            default_qdrant_url = os.environ.get("QDRANT_URL", "") or settings.qdrant_url
            default_qdrant_key = os.environ.get("QDRANT_API_KEY", "") or settings.qdrant_api_key
            
            qdrant_url = st.text_input(
                "Qdrant URL",
                value=default_qdrant_url,
                placeholder="https://xxx-xxx.aws.cloud.qdrant.io:6333",
                help="Your Qdrant Cloud cluster URL"
            )
            qdrant_api_key = st.text_input(
                "Qdrant API Key",
                type="password",
                value=default_qdrant_key,
                help="Your Qdrant API key"
            )
            st.session_state.qdrant_url = qdrant_url
            st.session_state.qdrant_api_key = qdrant_api_key
            
            if not qdrant_url or not qdrant_api_key:
                st.warning("⚠️ Get free Qdrant Cloud at: https://cloud.qdrant.io")
            
            # Test Qdrant connection
            if st.button("🔍 Test Qdrant Connection"):
                if qdrant_url and qdrant_api_key:
                    try:
                        test_store = create_vector_store("qdrant", url=qdrant_url, api_key=qdrant_api_key)
                        collections = test_store.list_collections()
                        st.success(f"✅ Connected! Found {len(collections)} collection(s)")
                    except Exception as e:
                        st.error(f"❌ Connection failed: {e}")
                else:
                    st.error("Please enter Qdrant URL and API Key")
        
        st.divider()
        
        # Get available collections based on provider
        available_collections = get_available_collections(vector_store_provider)
        st.session_state.available_collections = available_collections
        
        # Option 1: Use existing collection
        if available_collections:
            st.subheader("📚 Existing Collections")
            st.write(f"Found {len(available_collections)} collection(s)")
            
            selected_collection = st.selectbox(
                "Or select existing collection:",
                available_collections,
                key="collection_selector"
            )
            
            if st.button("📖 Load Existing Collection", type="secondary"):
                # Validate based on provider
                if llm_provider == "groq" and not groq_api_key:
                    st.error("Please enter your Groq API key")
                elif vector_store_provider == "qdrant" and (not st.session_state.get("qdrant_url") or not st.session_state.get("qdrant_api_key")):
                    st.error("Please enter Qdrant URL and API Key")
                else:
                    load_existing_collection(
                        groq_api_key, 
                        selected_collection,
                        llm_provider,
                        ollama_host if llm_provider == "ollama" else None,
                        vector_store_provider
                    )
            
            st.divider()
        
        # Option 2: Create new collection
        st.subheader("🆕 Create New Collection")
        
        # Dataset selection
        st.subheader("1. Dataset Selection")
        dataset_name = st.selectbox(
            "Choose Dataset",
            settings.ragbench_datasets,
            index=0
        )
        
        # Get dataset size dynamically
        if st.button("🔍 Check Dataset Size", key="check_size"):
            with st.spinner("Checking dataset size..."):
                try:
                    from datasets import load_dataset
                    
                    # Load dataset with download_mode to avoid cache issues
                    st.info(f"Fetching dataset info for '{dataset_name}'...")
                    ds = load_dataset(
                        "rungalileo/ragbench", 
                        dataset_name, 
                        split="train",
                        trust_remote_code=True,
                        download_mode="force_redownload"  # Force fresh download to avoid cache corruption
                    )
                    dataset_size = len(ds)
                    
                    st.session_state.dataset_size = dataset_size
                    st.session_state.current_dataset = dataset_name
                    st.success(f"✅ Dataset '{dataset_name}' has {dataset_size:,} samples available")
                except Exception as e:
                    st.error(f"❌ Error: {str(e)}")
                    st.exception(e)
                    st.warning(f"Could not determine dataset size. Using default of 10,000.")
                    st.session_state.dataset_size = 10000
                    st.session_state.current_dataset = dataset_name
        
        # Use stored dataset size or default
        max_samples_available = st.session_state.get('dataset_size', 10000)
        
        st.caption(f"Max available samples: {max_samples_available:,}")
        
        num_samples = st.slider(
            "Number of samples",
            min_value=10,
            max_value=max_samples_available,
            value=min(100, max_samples_available),
            step=50 if max_samples_available > 1000 else 10,
            help="Adjust slider to select number of samples"
        )
        
        load_all_samples = st.checkbox(
            "Load all available samples",
            value=False,
            help="Override slider and load entire dataset"
        )
        
        st.divider()
        
        # Chunking strategy
        st.subheader("2. Chunking Strategy")
        chunking_strategy = st.selectbox(
            "Choose Chunking Strategy",
            settings.chunking_strategies,
            index=0
        )
        
        chunk_size = st.slider(
            "Chunk Size",
            min_value=256,
            max_value=1024,
            value=512,
            step=128
        )
        
        overlap = st.slider(
            "Overlap",
            min_value=0,
            max_value=200,
            value=50,
            step=10
        )
        
        st.divider()
        
        # Embedding model
        st.subheader("3. Embedding Model")
        embedding_model = st.selectbox(
            "Choose Embedding Model",
            settings.embedding_models,
            index=0
        )
        
        st.divider()
        
        # LLM model selection for new collection
        st.subheader("4. LLM Model")
        if llm_provider == "groq":
            llm_model = st.selectbox(
                "Choose Groq LLM",
                settings.llm_models,
                index=1
            )
        else:
            llm_model = st.selectbox(
                "Choose Ollama Model",
                settings.ollama_models,
                index=settings.ollama_models.index(st.session_state.ollama_model) if st.session_state.ollama_model in settings.ollama_models else 0,
                key="llm_model_ollama"
            )
        
        st.divider()
        
        # Load data button
        if st.button("🚀 Load Data & Create Collection", type="primary"):
            # Validate based on provider
            if llm_provider == "groq" and not groq_api_key:
                st.error("Please enter your Groq API key")
            elif vector_store_provider == "qdrant" and (not st.session_state.get("qdrant_url") or not st.session_state.get("qdrant_api_key")):
                st.error("Please enter Qdrant URL and API Key")
            else:
                # Use None for num_samples if loading all data
                samples_to_load = None if load_all_samples else num_samples
                load_and_create_collection(
                    groq_api_key,
                    dataset_name,
                    samples_to_load,
                    chunking_strategy,
                    chunk_size,
                    overlap,
                    embedding_model,
                    llm_model,
                    llm_provider,
                    ollama_host if llm_provider == "ollama" else None,
                    vector_store_provider
                )
    
    # Main content area
    if not st.session_state.collection_loaded:
        st.info("👈 Please configure and load a dataset from the sidebar to begin")
        
        # Show instructions
        with st.expander("📖 How to Use", expanded=True):
            st.markdown("""
            1. **Enter your Groq API Key** in the sidebar
            2. **Select a dataset** from RAG Bench
            3. **Choose a chunking strategy** (dense, sparse, hybrid, re-ranking)
            4. **Select an embedding model** for document vectorization
            5. **Choose an LLM model** for response generation
            6. **Click "Load Data & Create Collection"** to initialize
            7. **Start chatting** in the chat interface
            8. **View retrieved documents** and evaluation metrics
            9. **Run TRACE evaluation** on test data
            """)
        
        # Show available options
        col1, col2 = st.columns(2)
        
        with col1:
            st.subheader("📊 Available Datasets")
            for ds in settings.ragbench_datasets:
                st.markdown(f"- {ds}")
        
        with col2:
            st.subheader("🤖 Available Models")
            st.markdown("**Embedding Models:**")
            for em in settings.embedding_models:
                st.markdown(f"- {em}")
            
            st.markdown("**LLM Models:**")
            for lm in settings.llm_models:
                st.markdown(f"- {lm}")
    
    else:
        # Create tabs for different functionalities
        tab1, tab2, tab3 = st.tabs(["💬 Chat", "📊 Evaluation", "📜 History"])
        
        with tab1:
            chat_interface()
        
        with tab2:
            evaluation_interface()
        
        with tab3:
            history_interface()


def load_existing_collection(api_key: str, collection_name: str, llm_provider: str = "groq", ollama_host: str = None, vector_store_provider: str = "chroma"):
    """Load an existing collection from vector store."""
    with st.spinner(f"Loading collection '{collection_name}'..."):
        try:
            # Initialize vector store based on provider
            if vector_store_provider == "qdrant":
                qdrant_url = st.session_state.get("qdrant_url") or settings.qdrant_url
                qdrant_api_key = st.session_state.get("qdrant_api_key") or settings.qdrant_api_key
                vector_store = create_vector_store("qdrant", url=qdrant_url, api_key=qdrant_api_key)
            else:
                vector_store = ChromaDBManager(settings.chroma_persist_directory)
            
            vector_store.get_collection(collection_name)
            
            # Extract dataset name from collection name (format: dataset_name_strategy_model)
            # Try to find which dataset this collection is based on
            dataset_name = None
            for ds in settings.ragbench_datasets:
                if collection_name.startswith(ds.replace("-", "_")):
                    dataset_name = ds
                    break
            
            if not dataset_name:
                dataset_name = collection_name.split("_")[0]  # Fallback: use first part
            
            # Prompt for LLM selection based on provider
            if llm_provider == "groq":
                st.session_state.current_llm = st.selectbox(
                    "Select Groq LLM for this collection:",
                    settings.llm_models,
                    key=f"llm_selector_{collection_name}"
                )
            else:
                st.session_state.current_llm = st.selectbox(
                    "Select Ollama Model for this collection:",
                    settings.ollama_models,
                    key=f"ollama_selector_{collection_name}"
                )
            
            # Initialize LLM client based on provider
            st.info(f"Initializing LLM client ({llm_provider})...")
            llm_client = create_llm_client(
                provider=llm_provider,
                api_key=api_key,
                api_keys=settings.groq_api_keys if settings.groq_api_keys else None,
                model_name=st.session_state.current_llm,
                ollama_host=ollama_host or settings.ollama_host,
                max_rpm=settings.groq_rpm_limit,
                rate_limit_delay=settings.rate_limit_delay,
                max_retries=settings.max_retries,
                retry_delay=settings.retry_delay
            )
            
            # Create RAG pipeline with correct parameter names
            st.info("Creating RAG pipeline...")
            rag_pipeline = RAGPipeline(
                llm_client=llm_client,
                vector_store_manager=vector_store
            )
            
            # Store in session state
            st.session_state.vector_store = vector_store
            st.session_state.rag_pipeline = rag_pipeline
            st.session_state.collection_loaded = True
            st.session_state.current_collection = collection_name
            st.session_state.selected_collection = collection_name
            st.session_state.groq_api_key = api_key
            st.session_state.dataset_name = dataset_name
            st.session_state.collection_name = collection_name
            st.session_state.llm_provider = llm_provider
            
            # Extract embedding model from collection metadata or vector store
            embedding_model_name = None
            
            # For ChromaDB: check collection metadata
            if hasattr(vector_store, 'current_collection') and vector_store.current_collection:
                if hasattr(vector_store.current_collection, 'metadata'):
                    collection_metadata = vector_store.current_collection.metadata
                    if collection_metadata and "embedding_model" in collection_metadata:
                        embedding_model_name = collection_metadata["embedding_model"]
            
            # For Qdrant or fallback: check if embedding_model was loaded on the vector store
            if not embedding_model_name and hasattr(vector_store, 'embedding_model') and vector_store.embedding_model:
                if hasattr(vector_store.embedding_model, 'model_name'):
                    embedding_model_name = vector_store.embedding_model.model_name
            
            # Set session state
            if embedding_model_name:
                st.session_state.embedding_model = embedding_model_name
            else:
                st.session_state.embedding_model = None
            
            # Display system prompt and model info
            provider_icon = "☁️" if llm_provider == "groq" else "🖥️"
            st.success(f"✅ Collection '{collection_name}' loaded successfully! {provider_icon} Using {llm_provider.upper()}")
            
            with st.expander("🤖 Model & System Prompt Information", expanded=False):
                col1, col2 = st.columns(2)
                with col1:
                    st.write(f"**Provider:** {provider_icon} {llm_provider.upper()}")
                    st.write(f"**Model:** {st.session_state.current_llm}")
                    st.write(f"**Collection:** {collection_name}")
                    st.write(f"**Dataset:** {dataset_name}")
                with col2:
                    st.write(f"**Temperature:** 0.0")
                    st.write(f"**Max Tokens:** 2048")
                    if llm_provider == "groq":
                        st.write(f"**Rate Limit:** {settings.groq_rpm_limit} RPM")
                    else:
                        st.write(f"**Rate Limit:** ✅ Unlimited (Local)")
                
                st.markdown("#### System Prompt")
                st.info("""
You are a Fact-Checking and Citation Specialist. Your task is to perform a rigorous audit of a response against provided documents to determine its accuracy, relevance, and level of support.

**Task:**
1. Analyze the provided documents and identify information relevant to the user's question
2. Evaluate the response sentence-by-sentence
3. Verify each response sentence maps to supporting document sentences
4. Identify which document sentences were actually used in the response
                """)
            
            st.rerun()
            
        except Exception as e:
            st.error(f"Error loading collection: {str(e)}")
            st.exception(e)


def load_and_create_collection(
    api_key: str,
    dataset_name: str,
    num_samples: Optional[int],
    chunking_strategy: str,
    chunk_size: int,
    overlap: int,
    embedding_model: str,
    llm_model: str,
    llm_provider: str = "groq",
    ollama_host: str = None,
    vector_store_provider: str = "chroma"
):
    """Load dataset and create vector collection."""
    with st.spinner("Loading dataset and creating collection..."):
        try:
            # Initialize dataset loader
            loader = RAGBenchLoader()
            
            # Load dataset
            if num_samples is None:
                st.info(f"Loading {dataset_name} dataset (all available samples)...")
            else:
                st.info(f"Loading {dataset_name} dataset ({num_samples} samples)...")
            dataset = loader.load_dataset(dataset_name, split="train", max_samples=num_samples)
            
            if not dataset:
                st.error("Failed to load dataset")
                return
            
            # Initialize vector store based on provider
            st.info(f"Initializing vector store ({vector_store_provider})...")
            if vector_store_provider == "qdrant":
                qdrant_url = st.session_state.get("qdrant_url") or settings.qdrant_url
                qdrant_api_key = st.session_state.get("qdrant_api_key") or settings.qdrant_api_key
                vector_store = create_vector_store("qdrant", url=qdrant_url, api_key=qdrant_api_key)
            else:
                vector_store = ChromaDBManager(settings.chroma_persist_directory)
            
            # Create collection name
            collection_name = f"{dataset_name}_{chunking_strategy}_{embedding_model.split('/')[-1]}"
            collection_name = collection_name.replace("-", "_").replace(".", "_")
            
            # Delete existing collection with same name (if exists)
            existing_collections = vector_store.list_collections()
            if collection_name in existing_collections:
                st.warning(f"Collection '{collection_name}' already exists. Deleting and recreating...")
                vector_store.delete_collection(collection_name)
                st.info("Old collection deleted. Creating new one...")
            
            # Load data into collection
            st.info(f"Creating collection with {chunking_strategy} chunking...")
            vector_store.load_dataset_into_collection(
                collection_name=collection_name,
                embedding_model_name=embedding_model,
                chunking_strategy=chunking_strategy,
                dataset_data=dataset,
                chunk_size=chunk_size,
                overlap=overlap
            )
            
            # Initialize LLM client based on provider
            st.info(f"Initializing LLM client ({llm_provider})...")
            llm_client = create_llm_client(
                provider=llm_provider,
                api_key=api_key,
                api_keys=settings.groq_api_keys if settings.groq_api_keys else None,
                model_name=llm_model,
                ollama_host=ollama_host or settings.ollama_host,
                max_rpm=settings.groq_rpm_limit,
                rate_limit_delay=settings.rate_limit_delay,
                max_retries=settings.max_retries,
                retry_delay=settings.retry_delay
            )
            
            # Create RAG pipeline with correct parameter names
            rag_pipeline = RAGPipeline(
                llm_client=llm_client,
                vector_store_manager=vector_store
            )
            
            # Store in session state
            st.session_state.vector_store = vector_store
            st.session_state.rag_pipeline = rag_pipeline
            st.session_state.collection_loaded = True
            st.session_state.current_collection = collection_name
            st.session_state.dataset_name = dataset_name
            st.session_state.dataset = dataset
            st.session_state.collection_name = collection_name
            st.session_state.embedding_model = embedding_model
            st.session_state.groq_api_key = api_key
            st.session_state.llm_provider = llm_provider
            st.session_state.vector_store_provider = vector_store_provider
            
            provider_icon = "☁️" if llm_provider == "groq" else "🖥️"
            vs_icon = "☁️" if vector_store_provider == "qdrant" else "💾"
            st.success(f"✅ Collection '{collection_name}' created successfully! {provider_icon} Using {llm_provider.upper()}")
            st.rerun()
            
        except Exception as e:
            st.error(f"Error: {str(e)}")


def chat_interface():
    """Chat interface tab."""
    st.subheader("💬 Chat Interface")
    
    # Check if collection is loaded
    if not st.session_state.collection_loaded:
        st.warning("⚠️ No data loaded. Please use the configuration panel to load a dataset and create a collection.")
        st.info("""
        Steps:
        1. Select a dataset from the dropdown
        2. Click "Load Data & Create Collection" button
        3. Wait for the collection to be created
        4. Then you can start chatting
        """)
        return
    
    # Display collection info and LLM selector
    col1, col2, col3 = st.columns([2, 2, 1])
    with col1:
        provider_icon = "☁️" if st.session_state.get("llm_provider", "groq") == "groq" else "🖥️"
        st.info(f"📚 Collection: {st.session_state.current_collection} | {provider_icon} {st.session_state.get('llm_provider', 'groq').upper()}")
    
    with col2:
        # LLM selector for chat - based on provider
        current_provider = st.session_state.get("llm_provider", "groq")
        if current_provider == "groq":
            model_options = settings.llm_models
            try:
                current_index = settings.llm_models.index(st.session_state.current_llm)
            except ValueError:
                current_index = 0
        else:
            model_options = settings.ollama_models
            try:
                current_index = settings.ollama_models.index(st.session_state.current_llm)
            except ValueError:
                current_index = 0
        
        selected_llm = st.selectbox(
            f"Select {'Groq' if current_provider == 'groq' else 'Ollama'} Model for chat:",
            model_options,
            index=current_index,
            key="chat_llm_selector"
        )
        
        if selected_llm != st.session_state.current_llm:
            st.session_state.current_llm = selected_llm
            # Recreate LLM client with new model
            llm_client = create_llm_client(
                provider=current_provider,
                api_key=st.session_state.groq_api_key if "groq_api_key" in st.session_state else "",
                api_keys=settings.groq_api_keys if settings.groq_api_keys else None,
                model_name=selected_llm,
                ollama_host=settings.ollama_host,
                max_rpm=settings.groq_rpm_limit,
                rate_limit_delay=settings.rate_limit_delay
            )
            st.session_state.rag_pipeline.llm = llm_client
    
    with col3:
        if st.button("🗑️ Clear History"):
            st.session_state.chat_history = []
            st.session_state.rag_pipeline.clear_history()
            st.rerun()
    
    # Show system prompt info in expandable section
    with st.expander("🤖 System Prompt & Model Info", expanded=False):
        current_provider = st.session_state.get("llm_provider", "groq")
        col1, col2 = st.columns(2)
        with col1:
            provider_icon = "☁️" if current_provider == "groq" else "🖥️"
            st.write(f"**Provider:** {provider_icon} {current_provider.upper()}")
            st.write(f"**LLM Model:** {st.session_state.current_llm}")
            st.write(f"**Temperature:** 0.0")
            st.write(f"**Max Tokens:** 2048")
        with col2:
            st.write(f"**Collection:** {st.session_state.current_collection}")
            st.write(f"**Dataset:** {st.session_state.get('dataset_name', 'N/A')}")
            if current_provider == "groq":
                st.write(f"**Rate Limit:** {settings.groq_rpm_limit} RPM")
            else:
                st.write(f"**Rate Limit:** ✅ Unlimited (Local)")
        
        st.markdown("#### System Prompt Being Used")
        system_prompt = """You are a Fact-Checking and Citation Specialist. Your task is to perform a rigorous audit of a response against provided documents to determine its accuracy, relevance, and level of support.

**TASK OVERVIEW**
1. **Analyze Documents**: Review the provided documents and identify information relevant to the user's question.
2. **Evaluate Response**: Review the provided answer sentence-by-sentence.
3. **Verify Support**: Map each answer sentence to specific supporting sentences in the documents.
4. **Identify Utilization**: Determine which document sentences were actually used (directly or implicitly) to form the answer."""
        st.info(system_prompt)
    
    # Chat container
    chat_container = st.container()
    
    # Display chat history
    with chat_container:
        for chat_idx, entry in enumerate(st.session_state.chat_history):
            # User message
            with st.chat_message("user"):
                st.write(entry["query"])
            
            # Assistant message
            with st.chat_message("assistant"):
                st.write(entry["response"])
                
                # Show retrieved documents in expander
                with st.expander("📄 Retrieved Documents"):
                    for doc_idx, doc in enumerate(entry["retrieved_documents"]):
                        st.markdown(f"**Document {doc_idx+1}** (Distance: {doc.get('distance', 'N/A'):.4f})")
                        st.text_area(
                            f"doc_{chat_idx}_{doc_idx}",
                            value=doc["document"],
                            height=100,
                            key=f"doc_area_{chat_idx}_{doc_idx}",
                            label_visibility="collapsed"
                        )
                        if doc.get("metadata"):
                            st.caption(f"Metadata: {doc['metadata']}")
    
    # Chat input
    query = st.chat_input("Ask a question...")
    
    if query:
        # Check if collection exists
        if not st.session_state.rag_pipeline or not st.session_state.rag_pipeline.vector_store.current_collection:
            st.error("❌ No data loaded. Please load a dataset first using the configuration panel.")
            st.stop()
        
        # Add user message
        with chat_container:
            with st.chat_message("user"):
                st.write(query)
        
        # Generate response
        with st.spinner("Generating response..."):
            try:
                result = st.session_state.rag_pipeline.query(query)
            except Exception as e:
                st.error(f"❌ Error querying: {str(e)}")
                st.info("Please load a dataset and create a collection first.")
                st.stop()
        
        # Add assistant message
        with chat_container:
            with st.chat_message("assistant"):
                st.write(result["response"])
                
                # Show retrieved documents
                with st.expander("📄 Retrieved Documents"):
                    for doc_idx, doc in enumerate(result["retrieved_documents"]):
                        st.markdown(f"**Document {doc_idx+1}** (Distance: {doc.get('distance', 'N/A'):.4f})")
                        st.text_area(
                            f"doc_current_{doc_idx}",
                            value=doc["document"],
                            height=100,
                            key=f"doc_current_area_{doc_idx}",
                            label_visibility="collapsed"
                        )
                        if doc.get("metadata"):
                            st.caption(f"Metadata: {doc['metadata']}")
        
        # Store in history
        st.session_state.chat_history.append(result)
        st.rerun()


def evaluation_interface():
    """Evaluation interface tab."""
    st.subheader("📊 RAG Evaluation")
    
    # Check if collection is loaded
    if not st.session_state.collection_loaded:
        st.warning("⚠️ No data loaded. Please load a collection first.")
        return
    
    # Evaluation method selector
    eval_col1, eval_col2 = st.columns([2, 1])
    with eval_col1:
        evaluation_method = st.radio(
            "Evaluation Method:",
            options=["TRACE (Heuristic)", "GPT Labeling (LLM-based)", "Hybrid (Both)"],
            horizontal=True,
            help="TRACE is fast (no LLM). GPT Labeling is accurate but slower (requires LLM calls)."
        )
    
    # Map UI labels to method IDs
    method_map = {
        "TRACE (Heuristic)": "trace",
        "GPT Labeling (LLM-based)": "gpt_labeling",
        "Hybrid (Both)": "hybrid"
    }
    selected_method = method_map[evaluation_method]
    
    # LLM selector for evaluation
    current_provider = st.session_state.get("llm_provider", "groq")
    col1, col2 = st.columns([3, 1])
    with col1:
        # Show provider-specific models
        if current_provider == "groq":
            model_options = settings.llm_models
            try:
                current_index = settings.llm_models.index(st.session_state.current_llm)
            except ValueError:
                current_index = 0
        else:
            model_options = settings.ollama_models
            try:
                current_index = settings.ollama_models.index(st.session_state.current_llm)
            except ValueError:
                current_index = 0
        
        selected_llm = st.selectbox(
            f"Select {'Groq' if current_provider == 'groq' else 'Ollama'} Model for evaluation:",
            model_options,
            index=current_index,
            key="eval_llm_selector"
        )
        
        # Show provider info
        provider_icon = "☁️" if current_provider == "groq" else "🖥️"
        if current_provider == "ollama":
            st.caption(f"{provider_icon} Using local Ollama - **No rate limits!** Fast evaluation possible.")
        else:
            st.caption(f"{provider_icon} Using Groq API - Rate limited to {settings.groq_rpm_limit} RPM")
    
    # Show method description
    method_descriptions = {
        "trace": """
        **TRACE Heuristic Method** (Fast, Rule-Based)
        - Utilization: How well the system uses retrieved documents
        - Relevance: Relevance of retrieved documents to the query
        - Adherence: How well the response adheres to the retrieved context
        - Completeness: How complete the response is in answering the query
        - ⚡ Speed: ~100ms per evaluation
        - 💰 Cost: Free (no API calls)
        """,
        "gpt_labeling": """
        **GPT Labeling Method** (Accurate, LLM-based)
        - Uses sentence-level LLM analysis (from RAGBench paper)
        - Context Relevance: Fraction of context relevant to query
        - Context Utilization: Fraction of relevant context used
        - Completeness: Fraction of relevant info covered
        - Adherence: Response supported by context (no hallucinations)
        - ⏱️ Speed: ~2-5 seconds per evaluation
        - 💰 Cost: ~$0.002-0.01 per evaluation
        """,
        "hybrid": """
        **Hybrid Method** (Comprehensive)
        - Runs both TRACE and GPT Labeling methods
        - Provides both fast and accurate evaluation metrics
        - Best for detailed analysis
        - ⏱️ Speed: ~3-6 seconds per evaluation
        - 💰 Cost: Same as GPT Labeling
        """
    }
    
    st.markdown(method_descriptions[selected_method])
    
    # Get maximum test samples available for current dataset
    try:
        loader = RAGBenchLoader()
        max_test_samples = loader.get_test_data_size(st.session_state.dataset_name)
        st.caption(f"📊 Available test samples: {max_test_samples:,}")
    except Exception as e:
        max_test_samples = 100
        st.caption(f"Available test samples: ~{max_test_samples} (estimated)")
    
    # Ensure min and max are reasonable
    max_test_samples = max(5, min(max_test_samples, 500))  # Cap at 500 for performance
    
    num_test_samples = st.slider(
        "Number of test samples",
        min_value=5,
        max_value=max_test_samples,
        value=min(10, max_test_samples),
        step=5
    )
    
    # Show warning for GPT labeling (API cost) - only for Groq
    if selected_method in ["gpt_labeling", "hybrid"]:
        current_provider = st.session_state.get("llm_provider", "groq")
        if current_provider == "groq":
            st.warning(f"⚠️ **{evaluation_method}** requires LLM API calls. This will incur costs and be slower due to rate limiting ({settings.groq_rpm_limit} RPM).")
        else:
            st.info(f"ℹ️ **{evaluation_method}** using local Ollama - **No rate limits!** Evaluation will be much faster.")
    
    # Background evaluation toggle
    run_in_background = st.checkbox(
        "🔄 Run in Background",
        value=False,
        help="Run evaluation in background. You can close the browser and check results later."
    )
    
    col_run, col_jobs = st.columns([1, 1])
    
    with col_run:
        if st.button("🔬 Run Evaluation", type="primary"):
            if run_in_background:
                # Start background job
                start_background_evaluation(
                    num_test_samples, 
                    selected_llm, 
                    selected_method
                )
            else:
                # Use selected LLM for evaluation
                run_evaluation(num_test_samples, selected_llm, selected_method)
    
    with col_jobs:
        if st.button("📋 View Background Jobs"):
            st.session_state.show_background_jobs = True
    
    # Show background jobs panel
    if st.session_state.get("show_background_jobs", False):
        display_background_jobs()
    
    # Display results
    if st.session_state.evaluation_results:
        results = st.session_state.evaluation_results
        
        st.success("✅ Evaluation Complete!")
        st.divider()
        st.markdown("## 📊 Evaluation Metrics")
        
        # Display aggregate scores - handle both TRACE and GPT Labeling metric names
        st.markdown("### Main Metrics")
        col1, col2, col3, col4, col5 = st.columns(5)
        
        # Determine which metrics are available
        utilization = results.get('utilization') or results.get('context_utilization', 0)
        relevance = results.get('relevance') or results.get('context_relevance', 0)
        adherence = results.get('adherence', 0)
        completeness = results.get('completeness', 0)
        average = results.get('average', 0)
        
        with col1:
            st.metric("📊 Utilization", f"{utilization:.3f}")
        with col2:
            st.metric("🎯 Relevance", f"{relevance:.3f}")
        with col3:
            st.metric("✅ Adherence", f"{adherence:.3f}")
        with col4:
            st.metric("📝 Completeness", f"{completeness:.3f}")
        with col5:
            st.metric("⭐ Average", f"{average:.3f}")
        
        # Detailed results summary - handle both metric types
        if "individual_scores" in results:
            with st.expander("📋 Summary Metrics by Query"):
                df = pd.DataFrame(results["individual_scores"])
                st.dataframe(df, use_container_width=True)
        
        # Detailed per-query results
        if "detailed_results" in results and results["detailed_results"]:
            with st.expander("🔍 Detailed Per-Query Analysis"):
                for query_result in results.get("detailed_results", []):
                    with st.expander(f"Query {query_result['query_id']}: {query_result['question'][:60]}..."):
                        st.markdown("### Question")
                        st.write(query_result['question'])
                        
                        st.markdown("### LLM Response")
                        st.write(query_result.get('llm_response', 'N/A'))
                        
                        st.markdown("### Retrieved Documents")
                        for doc_idx, doc in enumerate(query_result.get('retrieved_documents', []), 1):
                            with st.expander(f"📄 Document {doc_idx}"):
                                st.write(doc)
                        
                        if query_result.get('ground_truth'):
                            st.markdown("### Ground Truth")
                            st.write(query_result['ground_truth'])
                        
                        # Display metrics with correct labels based on method
                        metrics = query_result.get('metrics', {})
                        if metrics:
                            st.markdown("### Evaluation Metrics")
                            col1, col2, col3, col4, col5 = st.columns(5)
                            
                            # Get metric values (handle both TRACE and GPT names)
                            util_val = metrics.get('utilization') or metrics.get('context_utilization', 0)
                            rel_val = metrics.get('relevance') or metrics.get('context_relevance', 0)
                            adh_val = metrics.get('adherence', 0)
                            comp_val = metrics.get('completeness', 0)
                            avg_val = metrics.get('average', 0)
                            
                            with col1:
                                st.metric("Util", f"{util_val:.3f}")
                            with col2:
                                st.metric("Rel", f"{rel_val:.3f}")
                            with col3:
                                st.metric("Adh", f"{adh_val:.3f}")
                            with col4:
                                st.metric("Comp", f"{comp_val:.3f}")
                            with col5:
                                st.metric("Avg", f"{avg_val:.3f}")
        
        # For GPT Labeling and Hybrid methods, show additional metrics
        method = results.get("method", "")
        if "gpt_labeling" in method or "hybrid" in method:
            # Show RMSE aggregation metrics (consistency across evaluations)
            if "rmse_metrics" in results:
                st.markdown("### 📊 RMSE Aggregation (Metric Consistency)")
                rmse_data = results.get("rmse_metrics", {})
                
                rmse_cols = st.columns(4)
                with rmse_cols[0]:
                    rel_mean = rmse_data.get("context_relevance", {}).get("mean", 0)
                    rel_std = rmse_data.get("context_relevance", {}).get("std_dev", 0)
                    st.metric("Relevance", f"{rel_mean:.3f} ±{rel_std:.3f}", help="Mean and Std Dev")
                with rmse_cols[1]:
                    util_mean = rmse_data.get("context_utilization", {}).get("mean", 0)
                    util_std = rmse_data.get("context_utilization", {}).get("std_dev", 0)
                    st.metric("Utilization", f"{util_mean:.3f} ±{util_std:.3f}", help="Mean and Std Dev")
                with rmse_cols[2]:
                    comp_mean = rmse_data.get("completeness", {}).get("mean", 0)
                    comp_std = rmse_data.get("completeness", {}).get("std_dev", 0)
                    st.metric("Completeness", f"{comp_mean:.3f} ±{comp_std:.3f}", help="Mean and Std Dev")
                with rmse_cols[3]:
                    adh_mean = rmse_data.get("adherence", {}).get("mean", 0)
                    adh_std = rmse_data.get("adherence", {}).get("std_dev", 0)
                    st.metric("Adherence", f"{adh_mean:.3f} ±{adh_std:.3f}", help="Mean and Std Dev")
                
                # Show detailed RMSE statistics in expander
                with st.expander("See detailed RMSE aggregation statistics"):
                    for metric_name, metric_data in rmse_data.items():
                        st.write(f"**{metric_name}**")
                        col1, col2, col3, col4 = st.columns(4)
                        with col1:
                            st.write(f"Mean: {metric_data.get('mean', 0):.4f}")
                        with col2:
                            st.write(f"Std Dev: {metric_data.get('std_dev', 0):.4f}")
                        with col3:
                            st.write(f"Min: {metric_data.get('min', 0):.4f}")
                        with col4:
                            st.write(f"Max: {metric_data.get('max', 0):.4f}")
            
            # Show per-metric statistics if available
            if "per_metric_statistics" in results:
                st.markdown("### 📈 Per-Metric Statistics (Distribution)")
                stats_data = results.get("per_metric_statistics", {})
                
                stats_cols = st.columns(4)
                with stats_cols[0]:
                    rel_stats = stats_data.get("context_relevance", {})
                    st.metric("Relevance Mean", f"{rel_stats.get('mean', 0):.3f}", help=f"Median: {rel_stats.get('median', 0):.3f}")
                with stats_cols[1]:
                    util_stats = stats_data.get("context_utilization", {})
                    st.metric("Utilization Mean", f"{util_stats.get('mean', 0):.3f}", help=f"Median: {util_stats.get('median', 0):.3f}")
                with stats_cols[2]:
                    comp_stats = stats_data.get("completeness", {})
                    st.metric("Completeness Mean", f"{comp_stats.get('mean', 0):.3f}", help=f"Median: {comp_stats.get('median', 0):.3f}")
                with stats_cols[3]:
                    adh_stats = stats_data.get("adherence", {})
                    st.metric("Adherence Mean", f"{adh_stats.get('mean', 0):.3f}", help=f"Median: {adh_stats.get('median', 0):.3f}")
                
                # Show detailed statistics
                with st.expander("See detailed per-metric statistics"):
                    for metric_name, metric_stats in stats_data.items():
                        st.write(f"**{metric_name}**")
                        col1, col2 = st.columns(2)
                        with col1:
                            st.write(f"""
- Mean: {metric_stats.get('mean', 0):.4f}
- Median: {metric_stats.get('median', 0):.4f}
- Std Dev: {metric_stats.get('std_dev', 0):.4f}
- Min: {metric_stats.get('min', 0):.4f}
- Max: {metric_stats.get('max', 0):.4f}
                            """)
                        with col2:
                            st.write(f"""
- 25th percentile: {metric_stats.get('percentile_25', 0):.4f}
- 75th percentile: {metric_stats.get('percentile_75', 0):.4f}
- Perfect (>=0.95): {metric_stats.get('perfect_count', 0)}
- Poor (<0.3): {metric_stats.get('poor_count', 0)}
- Samples: {metric_stats.get('sample_count', 0)}
                            """)
            
            # Show RMSE vs RAGBench Ground Truth (per RAGBench paper requirement)
            if "rmse_vs_ground_truth" in results:
                st.markdown("### 📉 RMSE vs RAGBench Ground Truth")
                st.info("Compares predicted TRACE scores against original RAGBench dataset scores")
                rmse_gt = results.get("rmse_vs_ground_truth", {})
                per_metric_rmse = rmse_gt.get("per_metric_rmse", {})
                
                if per_metric_rmse:
                    rmse_gt_cols = st.columns(5)
                    with rmse_gt_cols[0]:
                        st.metric("Relevance RMSE", f"{per_metric_rmse.get('context_relevance', 0):.4f}", 
                                  delta=None, help="Lower is better (0 = perfect match)")
                    with rmse_gt_cols[1]:
                        st.metric("Utilization RMSE", f"{per_metric_rmse.get('context_utilization', 0):.4f}")
                    with rmse_gt_cols[2]:
                        st.metric("Completeness RMSE", f"{per_metric_rmse.get('completeness', 0):.4f}")
                    with rmse_gt_cols[3]:
                        st.metric("Adherence RMSE", f"{per_metric_rmse.get('adherence', 0):.4f}")
                    with rmse_gt_cols[4]:
                        agg_rmse = rmse_gt.get("aggregated_rmse", 0)
                        consistency = rmse_gt.get("consistency_score", 0)
                        st.metric("Aggregated RMSE", f"{agg_rmse:.4f}", 
                                  delta=f"Consistency: {consistency:.2%}", delta_color="normal")
            
            # Show AUCROC vs RAGBench Ground Truth (per RAGBench paper requirement)
            if "aucroc_vs_ground_truth" in results:
                st.markdown("### 📊 AUC-ROC vs RAGBench Ground Truth")
                st.info("Area Under ROC Curve comparing predicted vs ground truth binary classifications")
                auc_gt = results.get("aucroc_vs_ground_truth", {})
                
                if auc_gt:
                    auc_cols = st.columns(5)
                    with auc_cols[0]:
                        st.metric("Relevance AUC", f"{auc_gt.get('context_relevance', 0):.4f}",
                                  help="Higher is better (1.0 = perfect classification)")
                    with auc_cols[1]:
                        st.metric("Utilization AUC", f"{auc_gt.get('context_utilization', 0):.4f}")
                    with auc_cols[2]:
                        st.metric("Completeness AUC", f"{auc_gt.get('completeness', 0):.4f}")
                    with auc_cols[3]:
                        st.metric("Adherence AUC", f"{auc_gt.get('adherence', 0):.4f}")
                    with auc_cols[4]:
                        avg_auc = auc_gt.get("average", 0)
                        st.metric("Average AUC", f"{avg_auc:.4f}")
        
        # Download results
        st.divider()
        st.markdown("## 💾 Download Results")
        
        # Create a comprehensive download with all details
        download_data = {
            "evaluation_metadata": {
                "timestamp": datetime.now().isoformat(),
                "dataset": st.session_state.dataset_name,
                "method": results.get("evaluation_config", {}).get("evaluation_method", "gpt_labeling_prompts"),
                "total_samples": results.get("num_samples", 0),
                "embedding_model": st.session_state.embedding_model,
            },
            "aggregate_metrics": {
                "context_relevance": results.get("context_relevance") or results.get("relevance", 0),
                "context_utilization": results.get("context_utilization") or results.get("utilization", 0),
                "completeness": results.get("completeness", 0),
                "adherence": results.get("adherence", 0),
                "average": results.get("average", 0),
            },
            "rmse_metrics": results.get("rmse_metrics", {}),
            "per_metric_statistics": results.get("per_metric_statistics", {}),
            "rmse_vs_ground_truth": results.get("rmse_vs_ground_truth", {}),
            "aucroc_vs_ground_truth": results.get("aucroc_vs_ground_truth", {}),
            "detailed_results": results.get("detailed_results", [])
        }
        
        results_json = json.dumps(download_data, indent=2, default=str)
        
        col1, col2 = st.columns(2)
        with col1:
            st.download_button(
                label="📥 Download Complete Results (JSON)",
                data=results_json,
                file_name=f"evaluation_results_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
                mime="application/json",
                help="Download all evaluation results including metrics and per-query details"
            )
        with col2:
            st.download_button(
                label="📋 Download Metrics Only (JSON)",
                data=json.dumps(download_data["aggregate_metrics"], indent=2),
                file_name=f"evaluation_metrics_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
                mime="application/json",
                help="Download only the aggregate metrics"
            )


def start_background_evaluation(num_samples: int, selected_llm: str = None, method: str = "trace"):
    """Start evaluation in background thread.
    
    Args:
        num_samples: Number of test samples to evaluate
        selected_llm: LLM model to use for evaluation
        method: Evaluation method
    """
    if not BACKGROUND_EVAL_AVAILABLE:
        st.error("Background evaluation not available. Running in foreground.")
        run_evaluation(num_samples, selected_llm, method)
        return
    
    try:
        # Generate unique job ID
        job_id = f"{st.session_state.collection_name}_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{str(uuid.uuid4())[:8]}"
        
        # Create job record
        job = create_job(
            job_id=job_id,
            collection_name=st.session_state.collection_name,
            dataset_name=st.session_state.dataset_name,
            num_samples=num_samples,
            method=method,
            llm_model=selected_llm or st.session_state.current_llm,
            embedding_model=st.session_state.embedding_model,
            llm_provider=st.session_state.get("llm_provider", "groq")
        )
        
        # Start background thread
        start_background_job(
            job_id=job_id,
            rag_pipeline=st.session_state.rag_pipeline,
            vector_store=st.session_state.vector_store,
            dataset_name=st.session_state.dataset_name,
            num_samples=num_samples,
            method=method,
            llm_model=selected_llm or st.session_state.current_llm,
            embedding_model=st.session_state.embedding_model,
            llm_provider=st.session_state.get("llm_provider", "groq"),
            groq_api_key=st.session_state.get("groq_api_key"),
            groq_api_keys=settings.groq_api_keys if settings.groq_api_keys else None,
            ollama_host=settings.ollama_host
        )
        
        st.success(f"""
        ✅ **Background evaluation started!**
        
        - **Job ID:** `{job_id}`
        - **Method:** {method}
        - **Samples:** {num_samples}
        
        You can safely close this page. Check back later using "View Background Jobs" button.
        """)
        
        # Show the jobs panel
        st.session_state.show_background_jobs = True
        
    except Exception as e:
        st.error(f"Failed to start background evaluation: {str(e)}")


def display_background_jobs():
    """Display background jobs status panel."""
    st.divider()
    st.markdown("### 📋 Background Evaluation Jobs")
    
    if not BACKGROUND_EVAL_AVAILABLE:
        st.warning("Background evaluation module not available.")
        return
    
    # Refresh button
    if st.button("🔄 Refresh Status"):
        st.rerun()
    
    # Get all jobs
    jobs = list_jobs()
    
    if not jobs:
        st.info("No background jobs found.")
        if st.button("Close Panel"):
            st.session_state.show_background_jobs = False
            st.rerun()
        return
    
    # Display jobs
    for job in jobs[:10]:  # Show last 10 jobs
        job_id = job.get("job_id", "Unknown")
        status = job.get("status", "unknown")
        progress = job.get("progress", 0)
        
        # Status icon
        status_icons = {
            "pending": "⏳",
            "running": "🔄",
            "completed": "✅",
            "failed": "❌"
        }
        icon = status_icons.get(status, "❓")
        
        with st.expander(f"{icon} {job.get('collection_name', 'Unknown')} - {status.upper()} ({progress}%)", expanded=(status == "running")):
            col1, col2 = st.columns([2, 1])
            
            with col1:
                st.write(f"**Job ID:** `{job_id}`")
                st.write(f"**Method:** {job.get('method', 'Unknown')}")
                st.write(f"**Samples:** {job.get('num_samples', 0)}")
                st.write(f"**LLM:** {job.get('llm_model', 'Unknown')}")
                st.write(f"**Created:** {job.get('created_at', 'Unknown')}")
                
                if status == "running":
                    st.progress(progress / 100)
                    st.write(f"**Current Step:** {job.get('current_step', '')}")
                
                if status == "completed":
                    st.write(f"**Completed:** {job.get('completed_at', 'Unknown')}")
                    results_file = job.get('results_file')
                    if results_file:
                        st.write(f"**Results File:** `{results_file}`")
                    
                    # Load results into session state
                    if st.button(f"📊 Load Results", key=f"load_{job_id}"):
                        results = job.get("results")
                        if results:
                            st.session_state.evaluation_results = results
                            st.success("Results loaded! Scroll down to see metrics.")
                            st.rerun()
                
                if status == "failed":
                    st.error(f"**Error:** {job.get('error', 'Unknown error')}")
            
            with col2:
                # Show logs
                logs = job.get("logs", [])
                if logs:
                    st.write("**Recent Logs:**")
                    for log in logs[-5:]:  # Last 5 logs
                        st.caption(f"{log.get('message', '')}")
                
                # Delete button
                if st.button(f"🗑️ Delete", key=f"del_{job_id}"):
                    delete_job(job_id)
                    st.success(f"Job {job_id} deleted.")
                    st.rerun()
    
    if st.button("Close Panel"):
        st.session_state.show_background_jobs = False
        st.rerun()


def run_evaluation(num_samples: int, selected_llm: str = None, method: str = "trace"):
    """Run evaluation using selected method (TRACE, GPT Labeling, or Hybrid).
    
    Args:
        num_samples: Number of test samples to evaluate
        selected_llm: LLM model to use for evaluation
        method: Evaluation method ("trace", "gpt_labeling", or "hybrid")
    """
    with st.spinner(f"Running evaluation on {num_samples} samples..."):
        try:
            # Create logs container
            logs_container = st.container()
            logs_list = []
            
            # Display logs header once outside function
            logs_placeholder = st.empty()
            
            def add_log(message: str):
                """Add log message and update display."""
                logs_list.append(message)
                with logs_placeholder.container():
                    st.markdown("### 📋 Evaluation Logs:")
                    for log_msg in logs_list:
                        st.caption(log_msg)
            
            # Log evaluation start
            timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
            add_log(f"⏱️ Evaluation started at {timestamp}")
            add_log(f"📊 Dataset: {st.session_state.dataset_name}")
            add_log(f"📈 Total samples: {num_samples}")
            add_log(f"🤖 LLM Model: {selected_llm if selected_llm else st.session_state.current_llm}")
            add_log(f"🔗 Vector Store: {st.session_state.collection_name}")
            add_log(f"🧠 Embedding Model: {st.session_state.embedding_model}")
            
            # Map method names
            method_names = {
                "trace": "TRACE (Heuristic)",
                "gpt_labeling": "GPT Labeling (LLM-based)",
                "hybrid": "Hybrid (Both)"
            }
            add_log(f"🔬 Evaluation Method: {method_names.get(method, method)}")
            
            # Use selected LLM if provided - create with appropriate provider
            eval_llm_client = None
            original_llm = None
            current_provider = st.session_state.get("llm_provider", "groq")
            
            if selected_llm and selected_llm != st.session_state.current_llm:
                add_log(f"🔄 Switching LLM to {selected_llm} ({current_provider.upper()})...")
                groq_api_key = st.session_state.groq_api_key if "groq_api_key" in st.session_state else ""
                eval_llm_client = create_llm_client(
                    provider=current_provider,
                    api_key=groq_api_key,
                    api_keys=settings.groq_api_keys if settings.groq_api_keys else None,
                    model_name=selected_llm,
                    ollama_host=settings.ollama_host,
                    max_rpm=settings.groq_rpm_limit,
                    rate_limit_delay=settings.rate_limit_delay,
                    max_retries=settings.max_retries,
                    retry_delay=settings.retry_delay
                )
                # Temporarily replace LLM client
                original_llm = st.session_state.rag_pipeline.llm
                st.session_state.rag_pipeline.llm = eval_llm_client
            else:
                eval_llm_client = st.session_state.rag_pipeline.llm
            
            # Log provider info
            provider_icon = "☁️" if current_provider == "groq" else "🖥️"
            add_log(f"{provider_icon} LLM Provider: {current_provider.upper()}")
            
            # Get test data
            add_log("📥 Loading test data...")
            loader = RAGBenchLoader()
            test_data = loader.get_test_data(
                st.session_state.dataset_name,
                num_samples
            )
            add_log(f"✅ Loaded {len(test_data)} test samples")
            
            # Prepare test cases
            test_cases = []
            
            progress_bar = st.progress(0)
            status_text = st.empty()
            
            add_log("🔍 Processing samples...")
            for i, sample in enumerate(test_data):
                status_text.text(f"Processing sample {i+1}/{num_samples}")
                
                # Query the RAG system
                result = st.session_state.rag_pipeline.query(
                    sample["question"],
                    n_results=5
                )
                
                # Prepare test case
                test_cases.append({
                    "query": sample["question"],
                    "response": result["response"],
                    "retrieved_documents": [doc["document"] for doc in result["retrieved_documents"]],
                    "ground_truth": sample.get("answer", "")
                })
                
                # Update progress
                progress_bar.progress((i + 1) / num_samples)
                
                # Log every 10 samples
                if (i + 1) % 10 == 0 or (i + 1) == num_samples:
                    add_log(f"  ✓ Processed {i + 1}/{num_samples} samples")
            
            status_text.text(f"Running {method_names.get(method, method)} evaluation...")
            add_log(f"📊 Running evaluation using {method_names.get(method, method)}...")
            
            # Extract chunking and embedding metadata from session state
            # (These were stored when the collection was loaded/created)
            chunking_strategy = st.session_state.vector_store.chunking_strategy if st.session_state.vector_store else None
            embedding_model = st.session_state.embedding_model
            chunk_size = st.session_state.vector_store.chunk_size if st.session_state.vector_store else None
            chunk_overlap = st.session_state.vector_store.chunk_overlap if st.session_state.vector_store else None
            
            # Log retrieval configuration
            add_log(f"🔧 Retrieval Configuration:")
            add_log(f"  • Chunking Strategy: {chunking_strategy or 'Unknown'}")
            add_log(f"  • Chunk Size: {chunk_size or 'Unknown'}")
            add_log(f"  • Chunk Overlap: {chunk_overlap or 'Unknown'}")
            add_log(f"  • Embedding Model: {embedding_model or 'Unknown'}")
            
            # Import unified pipeline
            try:
                from evaluation_pipeline import UnifiedEvaluationPipeline
                
                # Run evaluation with metadata using unified pipeline
                pipeline = UnifiedEvaluationPipeline(
                    llm_client=eval_llm_client,
                    chunking_strategy=chunking_strategy,
                    embedding_model=embedding_model,
                    chunk_size=chunk_size,
                    chunk_overlap=chunk_overlap
                )
                
                # Run evaluation with selected method
                results = pipeline.evaluate_batch(test_cases, method=method)
                
            except ImportError:
                # Fallback to TRACE only if evaluation_pipeline module not available
                add_log("⚠️  evaluation_pipeline module not found, falling back to TRACE...")
                
                # Run evaluation with metadata using TRACE
                evaluator = TRACEEvaluator(
                    chunking_strategy=chunking_strategy,
                    embedding_model=embedding_model,
                    chunk_size=chunk_size,
                    chunk_overlap=chunk_overlap
                )
                results = evaluator.evaluate_batch(test_cases)
            
            st.session_state.evaluation_results = results
            
            # Log evaluation results summary
            add_log("✅ Evaluation completed successfully!")
            
            # Display appropriate metrics based on method
            if method == "trace":
                add_log(f"  • Utilization: {results.get('utilization', 0):.2%}")
                add_log(f"  • Relevance: {results.get('relevance', 0):.2%}")
                add_log(f"  • Adherence: {results.get('adherence', 0):.2%}")
                add_log(f"  • Completeness: {results.get('completeness', 0):.2%}")
                add_log(f"  • Average: {results.get('average', 0):.2%}")
            elif method == "gpt_labeling":
                if "context_relevance" in results:
                    add_log(f"  • Context Relevance: {results.get('context_relevance', 0):.2%}")
                    add_log(f"  • Context Utilization: {results.get('context_utilization', 0):.2%}")
                    add_log(f"  • Completeness: {results.get('completeness', 0):.2%}")
                    add_log(f"  • Adherence: {results.get('adherence', 0):.2%}")
                    add_log(f"  • Average: {results.get('average', 0):.2%}")
                # NEW: Display RMSE and AUCROC metrics if available
                if "rmse_metrics" in results:
                    add_log(f"📈 RMSE Metrics (vs ground truth):")
                    rmse_metrics = results.get("rmse_metrics", {})
                    add_log(f"    • Context Relevance RMSE: {rmse_metrics.get('relevance', 0):.4f}")
                    add_log(f"    • Context Utilization RMSE: {rmse_metrics.get('utilization', 0):.4f}")
                    add_log(f"    • Completeness RMSE: {rmse_metrics.get('completeness', 0):.4f}")
                    add_log(f"    • Adherence RMSE: {rmse_metrics.get('adherence', 0):.4f}")
                    add_log(f"    • Average RMSE: {rmse_metrics.get('average', 0):.4f}")
                if "auc_metrics" in results:
                    add_log(f"📊 AUCROC Metrics (binary classification):")
                    auc_metrics = results.get("auc_metrics", {})
                    add_log(f"    • Context Relevance AUCROC: {auc_metrics.get('relevance', 0):.4f}")
                    add_log(f"    • Context Utilization AUCROC: {auc_metrics.get('utilization', 0):.4f}")
                    add_log(f"    • Completeness AUCROC: {auc_metrics.get('completeness', 0):.4f}")
                    add_log(f"    • Adherence AUCROC: {auc_metrics.get('adherence', 0):.4f}")
                    add_log(f"    • Average AUCROC: {auc_metrics.get('average', 0):.4f}")
            elif method == "hybrid":
                add_log("  📊 TRACE Metrics:")
                trace_res = results.get("trace_results", {})
                add_log(f"    • Utilization: {trace_res.get('utilization', 0):.2%}")
                add_log(f"    • Relevance: {trace_res.get('relevance', 0):.2%}")
                add_log(f"    • Adherence: {trace_res.get('adherence', 0):.2%}")
                add_log(f"    • Completeness: {trace_res.get('completeness', 0):.2%}")
                add_log("  🧠 GPT Labeling Metrics:")
                gpt_res = results.get("gpt_results", {})
                add_log(f"    • Context Relevance: {gpt_res.get('context_relevance', 0):.2%}")
                add_log(f"    • Context Utilization: {gpt_res.get('context_utilization', 0):.2%}")
                add_log(f"    • Completeness: {gpt_res.get('completeness', 0):.2%}")
                add_log(f"    • Adherence: {gpt_res.get('adherence', 0):.2%}")
            
            # Restore original LLM if it was switched
            if selected_llm and selected_llm != st.session_state.current_llm and original_llm:
                st.session_state.rag_pipeline.llm = original_llm
                add_log(f"🔄 Restored original LLM")
            
            add_log(f"⏱️ Evaluation completed at {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
            
            # Auto-save evaluation results to JSON file
            try:
                collection_name = st.session_state.get("collection_name", "unknown_collection")
                timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
                filename = f"{collection_name}_{timestamp}_evaluation.json"
                
                # Prepare save data with metadata
                save_data = {
                    "collection_name": collection_name,
                    "timestamp": datetime.now().isoformat(),
                    "evaluation_method": method,
                    "num_samples": num_samples,
                    "embedding_model": st.session_state.get("embedding_model", "unknown"),
                    "llm_model": st.session_state.get("current_llm", "unknown"),
                    "results": results
                }
                
                # Save to file
                with open(filename, 'w', encoding='utf-8') as f:
                    json.dump(save_data, f, indent=2, default=str)
                
                add_log(f"💾 Results auto-saved to: {filename}")
                st.success(f"💾 Results saved to: {filename}")
            except Exception as save_error:
                add_log(f"⚠️ Could not auto-save results: {str(save_error)}")
            
        except Exception as e:
            st.error(f"Error during evaluation: {str(e)}")
            add_log(f"❌ Error: {str(e)}")


def history_interface():
    """History interface tab."""
    st.subheader("📜 Chat History")
    
    if not st.session_state.chat_history:
        st.info("No chat history yet. Start a conversation in the Chat tab!")
        return
    
    # Export history
    col1, col2 = st.columns([3, 1])
    with col2:
        history_json = json.dumps(st.session_state.chat_history, indent=2)
        st.download_button(
            label="💾 Export History",
            data=history_json,
            file_name=f"chat_history_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
            mime="application/json"
        )
    
    # Display history
    for i, entry in enumerate(st.session_state.chat_history):
        with st.expander(f"💬 Conversation {i+1}: {entry['query'][:50]}..."):
            st.markdown(f"**Query:** {entry['query']}")
            st.markdown(f"**Response:** {entry['response']}")
            st.markdown(f"**Timestamp:** {entry.get('timestamp', 'N/A')}")
            
            st.markdown("**Retrieved Documents:**")
            for j, doc in enumerate(entry["retrieved_documents"]):
                st.text_area(
                    f"Document {j+1}",
                    value=doc["document"],
                    height=100,
                    key=f"history_doc_{i}_{j}"
                )


if __name__ == "__main__":
    main()