vikramronavrsc's picture
Rename advanced-rag.py to advanced_rag.py
9955c93 verified
raw
history blame contribute delete
20.9 kB
# advanced_rag.py
import os
import tempfile
import shutil
import PyPDF2
import streamlit as st
import torch
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.llms import HuggingFaceHub
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain.chains import RetrievalQA, LLMChain
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain.docstore.document import Document
from langchain.prompts import PromptTemplate
import time
import psutil
import uuid
import atexit
from blockchain_utils_metamask import BlockchainManagerMetaMask
class AdvancedRAG:
def __init__(self,
llm_model_name="mistralai/Mistral-7B-Instruct-v0.2",
embedding_model_name="sentence-transformers/all-MiniLM-L6-v2",
chunk_size=1000,
chunk_overlap=200,
use_gpu=True,
use_blockchain=False,
contract_address=None):
"""
Initialize the advanced RAG system with multiple retrieval methods.
Args:
llm_model_name: The HuggingFace model for text generation
embedding_model_name: The HuggingFace model for embeddings
chunk_size: Size of document chunks
chunk_overlap: Overlap between chunks
use_gpu: Whether to use GPU acceleration
use_blockchain: Whether to enable blockchain verification
contract_address: Address of the deployed RAG Document Verifier contract
"""
self.llm_model_name = llm_model_name
self.embedding_model_name = embedding_model_name
self.use_gpu = use_gpu and torch.cuda.is_available()
self.use_blockchain = use_blockchain
# Device selection for embeddings
self.device = "cuda" if self.use_gpu else "cpu"
st.sidebar.info(f"Using device: {self.device}")
# Initialize text splitter
self.text_splitter = RecursiveCharacterTextSplitter(
chunk_size=chunk_size,
chunk_overlap=chunk_overlap,
length_function=len,
)
# Initialize embeddings model
self.embeddings = HuggingFaceEmbeddings(
model_name=embedding_model_name,
model_kwargs={"device": self.device}
)
# Initialize LLM using HuggingFaceHub
try:
# Use HF_TOKEN from environment variables
hf_token = os.environ.get("HF_TOKEN")
if not hf_token:
st.warning("No HuggingFace token found. Using model without authentication.")
self.llm = HuggingFaceHub(
repo_id=llm_model_name,
huggingfacehub_api_token=hf_token,
model_kwargs={"temperature": 0.7, "max_length": 1024}
)
except Exception as e:
st.error(f"Error initializing LLM: {str(e)}")
st.info("Trying to initialize with default model...")
# Fallback to a smaller model
self.llm = HuggingFaceHub(
repo_id="google/flan-t5-small",
model_kwargs={"temperature": 0.7, "max_length": 512}
)
# Initialize vector store
self.vector_store = None
self.documents_processed = 0
# Monitoring stats
self.processing_times = {}
# Initialize blockchain manager if enabled
self.blockchain = None
if use_blockchain:
try:
self.blockchain = BlockchainManagerMetaMask(
contract_address=contract_address
)
st.sidebar.success("Blockchain manager initialized. Please connect MetaMask to continue.")
except Exception as e:
st.sidebar.error(f"Failed to initialize blockchain manager: {str(e)}")
self.use_blockchain = False
def update_blockchain_connection(self, metamask_info):
"""Update blockchain connection with MetaMask info."""
if self.blockchain and metamask_info:
self.blockchain.update_connection(
is_connected=metamask_info.get("connected", False),
user_address=metamask_info.get("address"),
network_id=metamask_info.get("network_id")
)
return self.blockchain.is_connected
return False
def process_pdfs(self, pdf_files):
"""Process PDF files, create a vector store, and verify documents on blockchain."""
all_docs = []
with st.status("Processing PDF files...") as status:
# Create temporary directory for file storage
temp_dir = tempfile.mkdtemp()
st.session_state['temp_dir'] = temp_dir
# Monitor processing time and memory usage
start_time = time.time()
# Track memory before processing
mem_before = psutil.virtual_memory().used / (1024 * 1024 * 1024) # GB
# Process each PDF file
for i, pdf_file in enumerate(pdf_files):
try:
file_start_time = time.time()
# Save uploaded file to temp directory
pdf_path = os.path.join(temp_dir, pdf_file.name)
with open(pdf_path, "wb") as f:
f.write(pdf_file.getbuffer())
status.update(label=f"Processing {pdf_file.name} ({i+1}/{len(pdf_files)})...")
# Extract text from PDF
text = ""
with open(pdf_path, "rb") as f:
pdf = PyPDF2.PdfReader(f)
for page_num in range(len(pdf.pages)):
page = pdf.pages[page_num]
page_text = page.extract_text()
if page_text:
text += page_text + "\n\n"
# Create documents
docs = [Document(page_content=text, metadata={"source": pdf_file.name})]
# Split documents into chunks
split_docs = self.text_splitter.split_documents(docs)
all_docs.extend(split_docs)
# Verify document on blockchain if enabled and connected
if self.use_blockchain and self.blockchain and self.blockchain.is_connected:
try:
# Create a unique document ID
document_id = f"{pdf_file.name}_{uuid.uuid4().hex[:8]}"
# Verify document on blockchain
status.update(label=f"Verifying {pdf_file.name} on blockchain...")
verification = self.blockchain.verify_document(document_id, pdf_path)
if verification.get('status'): # Success
st.sidebar.success(f"✅ {pdf_file.name} verified on blockchain")
if 'tx_hash' in verification:
st.sidebar.info(f"Transaction: {verification['tx_hash'][:10]}...")
# Add blockchain metadata to documents
for doc in split_docs:
doc.metadata["blockchain"] = {
"verified": True,
"document_id": document_id,
"document_hash": verification.get("document_hash", ""),
"tx_hash": verification.get("tx_hash", ""),
"block_number": verification.get("block_number", 0)
}
else:
st.sidebar.warning(f"❌ Failed to verify {pdf_file.name} on blockchain")
if 'error' in verification:
st.sidebar.error(f"Error: {verification['error']}")
except Exception as e:
st.sidebar.error(f"Blockchain verification error: {str(e)}")
elif self.use_blockchain:
st.sidebar.warning("MetaMask not connected. Document not verified on blockchain.")
file_end_time = time.time()
processing_time = file_end_time - file_start_time
st.sidebar.success(f"Processed {pdf_file.name}: {len(split_docs)} chunks in {processing_time:.2f}s")
self.processing_times[pdf_file.name] = {
"chunks": len(split_docs),
"time": processing_time
}
except Exception as e:
st.sidebar.error(f"Error processing {pdf_file.name}: {str(e)}")
# Create vector store if we have documents
if all_docs:
status.update(label="Building vector index...")
try:
# Record the time taken to build the index
index_start_time = time.time()
# Create the vector store using FAISS
self.vector_store = FAISS.from_documents(all_docs, self.embeddings)
index_end_time = time.time()
index_time = index_end_time - index_start_time
# Track memory after processing
mem_after = psutil.virtual_memory().used / (1024 * 1024 * 1024) # GB
mem_used = mem_after - mem_before
total_time = time.time() - start_time
status.update(label=f"Completed processing {len(all_docs)} chunks in {total_time:.2f}s", state="complete")
# Save performance metrics
self.processing_times["index_building"] = index_time
self.processing_times["total_time"] = total_time
self.processing_times["memory_used_gb"] = mem_used
self.documents_processed = len(all_docs)
return True
except Exception as e:
st.error(f"Error creating vector store: {str(e)}")
status.update(label="Error creating vector store", state="error")
return False
else:
status.update(label="No content extracted from PDFs", state="error")
return False
def direct_retrieval(self, query):
"""
Direct retrieval method: simply returns the most relevant document chunks without LLM processing.
Args:
query: User's question
Returns:
dict: Results with raw document chunks
"""
if not self.vector_store:
return "Please upload and process PDF files first."
try:
# Start timing the query
query_start_time = time.time()
# Retrieve the most relevant documents
with st.status("Searching documents..."):
retriever = self.vector_store.as_retriever(search_kwargs={"k": 5})
docs = retriever.get_relevant_documents(query)
# Calculate query time
query_time = time.time() - query_start_time
# Format sources and create answer from sources directly
sources = []
answer = f"Here are the most relevant passages for your query:\n\n"
for i, doc in enumerate(docs):
# Extract blockchain verification info if available
blockchain_info = None
if "blockchain" in doc.metadata:
blockchain_info = {
"verified": doc.metadata["blockchain"]["verified"],
"document_id": doc.metadata["blockchain"]["document_id"],
"tx_hash": doc.metadata["blockchain"]["tx_hash"]
}
source_text = doc.page_content
answer += f"**Passage {i+1}** (from {doc.metadata.get('source', 'Unknown')}):\n{source_text}\n\n"
sources.append({
"content": source_text,
"source": doc.metadata.get("source", "Unknown"),
"blockchain": blockchain_info
})
# Log query to blockchain if enabled and connected
blockchain_log = None
if self.use_blockchain and self.blockchain and self.blockchain.is_connected:
try:
with st.status("Logging query to blockchain..."):
log_result = self.blockchain.log_query(query, answer)
if log_result.get("status"): # Success
blockchain_log = {
"logged": True,
"query_id": log_result.get("query_id", ""),
"tx_hash": log_result.get("tx_hash", ""),
"block_number": log_result.get("block_number", 0)
}
else:
st.error(f"Error logging to blockchain: {log_result.get('error', 'Unknown error')}")
except Exception as e:
st.error(f"Error logging to blockchain: {str(e)}")
return {
"answer": answer,
"sources": sources,
"query_time": query_time,
"blockchain_log": blockchain_log,
"method": "direct"
}
except Exception as e:
st.error(f"Error in direct retrieval: {str(e)}")
return f"Error: {str(e)}"
def enhanced_retrieval(self, query):
"""
Enhanced retrieval method: uses an LLM to process the retrieved documents and generate a comprehensive answer.
Args:
query: User's question
Returns:
dict: Results with LLM-enhanced answer
"""
if not self.vector_store:
return "Please upload and process PDF files first."
try:
# Custom prompt for advanced processing
prompt_template = """
You are an AI research assistant with expertise in analyzing and synthesizing information from documents.
Below are relevant passages from documents that might answer the user's question.
USER QUESTION: {question}
RELEVANT PASSAGES:
{context}
Based on ONLY these passages, provide a comprehensive, accurate and well-structured answer to the question.
Your answer should:
1. Directly address the user's question
2. Synthesize information from multiple passages when applicable
3. Be detailed, precise and factual
4. Include specific examples or evidence from the passages
5. Acknowledge any limitations or gaps in the provided information
If the information to answer the question is not present in the passages, clearly state: "I don't have enough information to answer this question based on the available documents."
ANSWER:
"""
PROMPT = PromptTemplate(
template=prompt_template,
input_variables=["context", "question"]
)
# Start timing the query
query_start_time = time.time()
# Create QA chain
retriever = self.vector_store.as_retriever(search_kwargs={"k": 5})
# Get documents first to track sources
with st.status("Retrieving relevant documents..."):
docs = retriever.get_relevant_documents(query)
# Format sources
sources = []
for i, doc in enumerate(docs):
# Extract blockchain verification info if available
blockchain_info = None
if "blockchain" in doc.metadata:
blockchain_info = {
"verified": doc.metadata["blockchain"]["verified"],
"document_id": doc.metadata["blockchain"]["document_id"],
"tx_hash": doc.metadata["blockchain"]["tx_hash"]
}
sources.append({
"content": doc.page_content[:300] + "..." if len(doc.page_content) > 300 else doc.page_content,
"source": doc.metadata.get("source", "Unknown"),
"blockchain": blockchain_info
})
# Create document chain
document_chain = create_stuff_documents_chain(self.llm, PROMPT)
# Generate answer
with st.status("Generating enhanced answer..."):
answer = document_chain.invoke({
"question": query,
"context": docs
})
# Calculate query time
query_time = time.time() - query_start_time
# Log query to blockchain if enabled and connected
blockchain_log = None
if self.use_blockchain and self.blockchain and self.blockchain.is_connected:
try:
with st.status("Logging query to blockchain..."):
log_result = self.blockchain.log_query(query, answer)
if log_result.get("status"): # Success
blockchain_log = {
"logged": True,
"query_id": log_result.get("query_id", ""),
"tx_hash": log_result.get("tx_hash", ""),
"block_number": log_result.get("block_number", 0)
}
else:
st.error(f"Error logging to blockchain: {log_result.get('error', 'Unknown error')}")
except Exception as e:
st.error(f"Error logging to blockchain: {str(e)}")
return {
"answer": answer,
"sources": sources,
"query_time": query_time,
"blockchain_log": blockchain_log,
"method": "enhanced"
}
except Exception as e:
st.error(f"Error in enhanced retrieval: {str(e)}")
return f"Error: {str(e)}"
def ask(self, query, method="enhanced"):
"""
Ask a question using the specified retrieval method.
Args:
query: User's question
method: Retrieval method ("direct" or "enhanced")
Returns:
dict: Results from the specified retrieval method
"""
if method == "direct":
return self.direct_retrieval(query)
else:
return self.enhanced_retrieval(query)
def get_performance_metrics(self):
"""Return performance metrics for the RAG system."""
if not self.processing_times:
return None
return {
"documents_processed": self.documents_processed,
"index_building_time": self.processing_times.get("index_building", 0),
"total_processing_time": self.processing_times.get("total_time", 0),
"memory_used_gb": self.processing_times.get("memory_used_gb", 0),
"device": self.device,
"embedding_model": self.embedding_model_name,
"blockchain_enabled": self.use_blockchain,
"blockchain_connected": self.blockchain.is_connected if self.blockchain else False
}