RAG-PDF-Chatbot / app4.py
arjunanand13's picture
Rename app.py to app4.py
05fb03d verified
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_huggingface 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 dataset with proper error handling"""
try:
if dataset_name == "squad":
dataset = load_dataset("squad_v2", split="validation")
samples = dataset.select(range(0, 1000, 100))[:num_samples]
self.test_samples = []
for sample in samples:
# Handle SQuAD format
answers = sample["answers"]
if answers["text"]: # Check if there are answers
self.test_samples.append({
"question": sample["question"],
"ground_truth": answers["text"][0],
"context": sample["context"]
})
elif dataset_name == "msmarco":
dataset = load_dataset("ms_marco", "v2.1", split="test") # Changed from dev to test
samples = dataset.select(range(0, 1000, 100))[:num_samples]
self.test_samples = []
for sample in samples:
if sample["answers"]: # Check if answers exist
self.test_samples.append({
"question": sample["query"],
"ground_truth": sample["answers"][0],
"context": sample["passages"]["passage_text"][0]
})
self.current_dataset = dataset_name
return {
"dataset": dataset_name,
"samples_loaded": len(self.test_samples),
"example_questions": [s["question"] for s in self.test_samples[:3]]
}
except Exception as e:
print(f"Error loading dataset: {str(e)}")
return {
"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()):
"""Initialize vector database with error handling"""
try:
if not list_file_obj:
return None, "No files uploaded. Please upload PDF documents first."
list_file_path = [x.name for x in list_file_obj if x is not None]
if not list_file_path:
return None, "No valid files found. Please upload PDF documents."
doc_splits = load_doc(list_file_path, splitting_strategy, chunk_size)
if not doc_splits:
return None, "No content extracted from documents."
vector_db = create_db(doc_splits, db_choice)
return vector_db, f"Database created successfully using {splitting_strategy} splitting and {db_choice} vector database!"
except Exception as e:
return None, f"Error creating database: {str(e)}"
def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
"""Initialize LLM chain with error handling"""
try:
if vector_db is None:
return None, "Please create vector database first."
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!"
except Exception as e:
return None, f"Error initializing LLM: {str(e)}"
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]
).then(
lambda x: gr.update(interactive=True) if x[0] is not None else gr.update(interactive=False),
inputs=[vector_db],
outputs=[init_llm_btn]
)
init_llm_btn.click(
initialize_llmchain,
inputs=[llm_choice, temperature, max_tokens, top_k, vector_db],
outputs=[qa_chain, llm_progress]
).then(
lambda x: gr.update(interactive=True) if x[0] is not None else gr.update(interactive=False),
inputs=[qa_chain],
outputs=[msg]
)
load_dataset_btn.click(
lambda x: evaluator.load_dataset(x),
inputs=[dataset_choice],
outputs=[dataset_info]
)
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()