app.py

import gradio as gr
import os
from typing import List, Dict
import numpy as np
from datasets import load_dataset
from langchain.text_splitter import (
    RecursiveCharacterTextSplitter,
    CharacterTextSplitter,
    TokenTextSplitter
)
from langchain_community.vectorstores import FAISS, Chroma, Qdrant
from langchain_community.document_loaders import PyPDFLoader
from langchain.chains import ConversationalRetrievalChain
from langchain_community.embeddings import HuggingFaceEmbeddings 
from langchain_community.llms import HuggingFaceEndpoint
from langchain.memory import ConversationBufferMemory
from sentence_transformers import SentenceTransformer, util
import torch
from ragas import evaluate
from ragas.metrics import (
    ContextRecall,
    AnswerRelevancy,
    Faithfulness,
    ContextPrecision
)
import pandas as pd

# Constants and setup
list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"]
list_llm_simple = [os.path.basename(llm) for llm in list_llm]
api_token = os.getenv("HF_TOKEN")

CHUNK_SIZES = {
    "small": {"recursive": 512, "fixed": 512, "token": 256},
    "medium": {"recursive": 1024, "fixed": 1024, "token": 512}
}

# Initialize sentence transformer for evaluation
sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')

class RAGEvaluator:
    def __init__(self):
        self.datasets = {
            "squad": "squad_v2",
            "msmarco": "ms_marco"
        }
        self.current_dataset = None
        self.test_samples = []
        
    def load_dataset(self, dataset_name: str, num_samples: int = 10):
        """Load a smaller subset of questions with proper error handling"""
        try:
            if dataset_name == "squad":
                dataset = load_dataset("squad_v2", split="validation")
                # Select diverse questions
                samples = dataset.select(range(0, 1000, 100))[:num_samples]
                
                self.test_samples = []
                for sample in samples:
                    # Check if answers exist and are not empty
                    if sample.get("answers") and isinstance(sample["answers"], dict) and sample["answers"].get("text"):
                        self.test_samples.append({
                            "question": sample["question"],
                            "ground_truth": sample["answers"]["text"][0],
                            "context": sample["context"]
                        })
                
            elif dataset_name == "msmarco":
                dataset = load_dataset("ms_marco", "v2.1", split="dev")
                samples = dataset.select(range(0, 1000, 100))[:num_samples]
                
                self.test_samples = []
                for sample in samples:
                    # Check for valid answers
                    if sample.get("answers") and sample["answers"]:
                        self.test_samples.append({
                            "question": sample["query"],
                            "ground_truth": sample["answers"][0],
                            "context": sample["passages"][0]["passage_text"] 
                                     if isinstance(sample["passages"], list) 
                                     else sample["passages"]["passage_text"][0]
                        })
            
            self.current_dataset = dataset_name
            
            # Return dataset info
            return {
                "dataset": dataset_name,
                "num_samples": len(self.test_samples),
                "sample_questions": [s["question"] for s in self.test_samples[:3]],
                "status": "success"
            }
            
        except Exception as e:
            print(f"Error loading dataset: {str(e)}")
            return {
                "dataset": dataset_name,
                "error": str(e),
                "status": "failed"
            }

    def evaluate_configuration(self, vector_db, qa_chain, splitting_strategy: str, chunk_size: str) -> Dict:
        """Evaluate with progress tracking and error handling"""
        if not self.test_samples:
            return {"error": "No dataset loaded"}
            
        results = []
        total_questions = len(self.test_samples)
        
        # Add progress tracking
        for i, sample in enumerate(self.test_samples):
            print(f"Evaluating question {i+1}/{total_questions}")
            
            try:
                response = qa_chain.invoke({
                    "question": sample["question"],
                    "chat_history": []
                })
                
                results.append({
                    "question": sample["question"],
                    "answer": response["answer"],
                    "contexts": [doc.page_content for doc in response["source_documents"]],
                    "ground_truths": [sample["ground_truth"]]
                })
            except Exception as e:
                print(f"Error processing question {i+1}: {str(e)}")
                continue
        
        if not results:
            return {
                "configuration": f"{splitting_strategy}_{chunk_size}",
                "error": "No successful evaluations",
                "questions_evaluated": 0
            }
            
        try:
            # Calculate RAGAS metrics
            eval_dataset = Dataset.from_list(results)
            metrics = [ContextRecall(), AnswerRelevancy(), Faithfulness(), ContextPrecision()]
            scores = evaluate(eval_dataset, metrics=metrics)
            
            return {
                "configuration": f"{splitting_strategy}_{chunk_size}",
                "questions_evaluated": len(results),
                "context_recall": float(scores['context_recall']),
                "answer_relevancy": float(scores['answer_relevancy']),
                "faithfulness": float(scores['faithfulness']),
                "context_precision": float(scores['context_precision']),
                "average_score": float(np.mean([
                    scores['context_recall'],
                    scores['answer_relevancy'],
                    scores['faithfulness'],
                    scores['context_precision']
                ]))
            }
        except Exception as e:
            return {
                "configuration": f"{splitting_strategy}_{chunk_size}",
                "error": str(e),
                "questions_evaluated": len(results)
            }

# Text splitting and database functions
def get_text_splitter(strategy: str, chunk_size: int = 1024, chunk_overlap: int = 64):
    splitters = {
        "recursive": RecursiveCharacterTextSplitter(
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap
        ),
        "fixed": CharacterTextSplitter(
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap
        ),
        "token": TokenTextSplitter(
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap
        )
    }
    return splitters.get(strategy)

def load_doc(list_file_path: List[str], splitting_strategy: str, chunk_size: str):
    chunk_size_value = CHUNK_SIZES[chunk_size][splitting_strategy]
    loaders = [PyPDFLoader(x) for x in list_file_path]
    pages = []
    for loader in loaders:
        pages.extend(loader.load())
    
    text_splitter = get_text_splitter(splitting_strategy, chunk_size_value)
    doc_splits = text_splitter.split_documents(pages)
    return doc_splits

def create_db(splits, db_choice: str = "faiss"):
    embeddings = HuggingFaceEmbeddings()
    db_creators = {
        "faiss": lambda: FAISS.from_documents(splits, embeddings),
        "chroma": lambda: Chroma.from_documents(splits, embeddings),
        "qdrant": lambda: Qdrant.from_documents(
            splits,
            embeddings,
            location=":memory:",
            collection_name="pdf_docs"
        )
    }
    return db_creators[db_choice]()

def initialize_database(list_file_obj, splitting_strategy, chunk_size, db_choice, progress=gr.Progress()):
    list_file_path = [x.name for x in list_file_obj if x is not None]
    doc_splits = load_doc(list_file_path, splitting_strategy, chunk_size)
    vector_db = create_db(doc_splits, db_choice)
    return vector_db, f"Database created using {splitting_strategy} splitting and {db_choice} vector database!"

def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    llm_model = list_llm[llm_choice]
    
    llm = HuggingFaceEndpoint(
        repo_id=llm_model,
        huggingfacehub_api_token=api_token,
        temperature=temperature,
        max_new_tokens=max_tokens,
        top_k=top_k
    )
    
    memory = ConversationBufferMemory(
        memory_key="chat_history",
        output_key='answer',
        return_messages=True
    )

    retriever = vector_db.as_retriever()
    qa_chain = ConversationalRetrievalChain.from_llm(
        llm,
        retriever=retriever,
        memory=memory,
        return_source_documents=True
    )
    return qa_chain, "LLM initialized successfully!"

def conversation(qa_chain, message, history):
    """Fixed conversation function returning all required outputs"""
    response = qa_chain.invoke({
        "question": message,
        "chat_history": [(hist[0], hist[1]) for hist in history]
    })
    
    response_answer = response["answer"]
    if "Helpful Answer:" in response_answer:
        response_answer = response_answer.split("Helpful Answer:")[-1]
    
    # Get source documents, ensure we have exactly 3
    sources = response["source_documents"][:3]
    source_contents = []
    source_pages = []
    
    # Process available sources
    for source in sources:
        source_contents.append(source.page_content.strip())
        source_pages.append(source.metadata.get("page", 0) + 1)
    
    # Pad with empty values if we have fewer than 3 sources
    while len(source_contents) < 3:
        source_contents.append("")
        source_pages.append(0)
    
    # Return all required outputs in correct order
    return (
        qa_chain,  # State
        gr.update(value=""),  # Clear message box
        history + [(message, response_answer)],  # Updated chat history
        source_contents[0],  # First source
        source_pages[0],    # First page
        source_contents[1],  # Second source
        source_pages[1],    # Second page
        source_contents[2],  # Third source
        source_pages[2]     # Third page
    )

def demo():
    evaluator = RAGEvaluator()
    
    with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky")) as demo:
        vector_db = gr.State()
        qa_chain = gr.State()
        
        gr.HTML("<center><h1>Enhanced RAG PDF Chatbot with Evaluation</h1></center>")
        
        with gr.Tabs():
            # Custom PDF Tab
            with gr.Tab("Custom PDF Chat"):
                with gr.Row():
                    with gr.Column(scale=86):
                        gr.Markdown("<b>Step 1 - Configure and Initialize RAG Pipeline</b>")
                        with gr.Row():
                            document = gr.Files(
                                height=300,
                                file_count="multiple",
                                file_types=["pdf"],
                                interactive=True,
                                label="Upload PDF documents"
                            )
                        
                        with gr.Row():
                            splitting_strategy = gr.Radio(
                                ["recursive", "fixed", "token"],
                                label="Text Splitting Strategy",
                                value="recursive"
                            )
                            db_choice = gr.Radio(
                                ["faiss", "chroma", "qdrant"],
                                label="Vector Database",
                                value="faiss"
                            )
                            chunk_size = gr.Radio(
                                ["small", "medium"],
                                label="Chunk Size",
                                value="medium"
                            )
                        
                        with gr.Row():
                            db_btn = gr.Button("Create vector database")
                            db_progress = gr.Textbox(
                                value="Not initialized",
                                show_label=False
                            )
                        
                        gr.Markdown("<b>Step 2 - Configure LLM</b>")
                        with gr.Row():
                            llm_choice = gr.Radio(
                                list_llm_simple,
                                label="Available LLMs",
                                value=list_llm_simple[0],
                                type="index"
                            )
                        
                        with gr.Row():
                            with gr.Accordion("LLM Parameters", open=False):
                                temperature = gr.Slider(
                                    minimum=0.01,
                                    maximum=1.0,
                                    value=0.5,
                                    step=0.1,
                                    label="Temperature"
                                )
                                max_tokens = gr.Slider(
                                    minimum=128,
                                    maximum=4096,
                                    value=2048,
                                    step=128,
                                    label="Max Tokens"
                                )
                                top_k = gr.Slider(
                                    minimum=1,
                                    maximum=10,
                                    value=3,
                                    step=1,
                                    label="Top K"
                                )
                        
                        with gr.Row():
                            init_llm_btn = gr.Button("Initialize LLM")
                            llm_progress = gr.Textbox(
                                value="Not initialized",
                                show_label=False
                            )
                    
                    with gr.Column(scale=200):
                        gr.Markdown("<b>Step 3 - Chat with Documents</b>")
                        chatbot = gr.Chatbot(height=505)
                        
                        with gr.Accordion("Source References", open=False):
                            with gr.Row():
                                source1 = gr.Textbox(label="Source 1", lines=2)
                                page1 = gr.Number(label="Page")
                            with gr.Row():
                                source2 = gr.Textbox(label="Source 2", lines=2)
                                page2 = gr.Number(label="Page")
                            with gr.Row():
                                source3 = gr.Textbox(label="Source 3", lines=2)
                                page3 = gr.Number(label="Page")
                        
                        with gr.Row():
                            msg = gr.Textbox(
                                placeholder="Ask a question",
                                show_label=False
                            )
                        with gr.Row():
                            submit_btn = gr.Button("Submit")
                            clear_btn = gr.ClearButton(
                                [msg, chatbot],
                                value="Clear Chat"
                            )

            # Evaluation Tab
            with gr.Tab("RAG Evaluation"):
                with gr.Row():
                    dataset_choice = gr.Dropdown(
                        choices=list(evaluator.datasets.keys()),
                        label="Select Evaluation Dataset",
                        value="squad"
                    )
                    load_dataset_btn = gr.Button("Load Dataset")
                
                with gr.Row():
                    dataset_info = gr.JSON(label="Dataset Information")
                
                with gr.Row():
                    eval_splitting_strategy = gr.Radio(
                        ["recursive", "fixed", "token"],
                        label="Text Splitting Strategy",
                        value="recursive"
                    )
                    eval_chunk_size = gr.Radio(
                        ["small", "medium"],
                        label="Chunk Size",
                        value="medium"
                    )
                
                with gr.Row():
                    evaluate_btn = gr.Button("Run Evaluation")
                    evaluation_results = gr.DataFrame(label="Evaluation Results")

        # Event handlers
        db_btn.click(
            initialize_database,
            inputs=[document, splitting_strategy, chunk_size, db_choice],
            outputs=[vector_db, db_progress]
        )
        
        init_llm_btn.click(
            initialize_llmchain,
            inputs=[llm_choice, temperature, max_tokens, top_k, vector_db],
            outputs=[qa_chain, llm_progress]
        )
        
        msg.submit(
            conversation,
            inputs=[qa_chain, msg, chatbot],
            outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
        )
        
        submit_btn.click(
            conversation,
            inputs=[qa_chain, msg, chatbot],
            outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
        )

        def load_dataset_handler(dataset_name):
            try:
                result = evaluator.load_dataset(dataset_name)
                if result.get("status") == "success":
                    return {
                        "dataset": result["dataset"],
                        "samples_loaded": result["num_samples"],
                        "example_questions": result["sample_questions"],
                        "status": "ready for evaluation"
                    }
                else:
                    return {
                        "error": result.get("error", "Unknown error occurred"),
                        "status": "failed to load dataset"
                    }
            except Exception as e:
                return {
                    "error": str(e),
                    "status": "failed to load dataset"
                }
        
        def run_evaluation(dataset_choice, splitting_strategy, chunk_size, vector_db, qa_chain):
            if not evaluator.current_dataset:
                return pd.DataFrame()
            
            results = evaluator.evaluate_configuration(
                vector_db=vector_db,
                qa_chain=qa_chain,
                splitting_strategy=splitting_strategy,
                chunk_size=chunk_size
            )
            
            return pd.DataFrame([results])

        load_dataset_btn.click(
            load_dataset_handler,
            inputs=[dataset_choice],
            outputs=[dataset_info]
        )
        
        evaluate_btn.click(
            run_evaluation,
            inputs=[
                dataset_choice,
                eval_splitting_strategy,
                eval_chunk_size,
                vector_db,
                qa_chain
            ],
            outputs=[evaluation_results]
        )

        # Clear button handlers
        clear_btn.click(
            lambda: [None, "", 0, "", 0, "", 0],
            outputs=[chatbot, source1, page1, source2, page2, source3, page3]
        )

    # Launch the demo
    demo.queue().launch(debug=True)

if __name__ == "__main__":
    demo()