Behind Every Question-Answer AI Is a Data Pipeline Built for Scale — Here's How to Build Your Own
2024-11-26 16:50:1 Author: hackernoon.com(查看原文) 阅读量:0 收藏
2024-11-26 16:50:1 Author: hackernoon.com(查看原文) 阅读量:0 收藏
In this article, we’ll explore how to build a data pipeline that continuously indexes document embeddings into a Redis vector database. This solution leverages various Google Cloud services, including Cloud Storage, CloudSQL, Cloud Composer (Airflow), and Cloud Run, along with LangChain, OpenAI, and Redis as the vector store. The indexed documents can then be utilized in an RAG based question-answering system.
Section 1: Solution Overview
- GCP Storage Bucket: Serves as the central repository for all document types, such as PDFs, text files, and HTML .
- Document Ingestion: Collects newly created documents and stores them in the GCP Storage bucket.
- Airflow (Cloud Composer) as the Orchestrator: Automates and manages the entire workflow on a daily basis. Tasks are implemented using Google Cloud Run Operators.
Airflow Tasks:
- Get documents from Confluence created in the last 24 hours.
- Get HTML content created from the organization’s CMS in the last 24 hours.
- Get daily process notifications from Splunk alerts (webhook).
- Get daily summaries from specific services in New Relic.
- Retrieve documents from additional sources, such as source code from git repo, and use
Repopack to bundle each repository’s code (*.py , *.java) into a single txt file. - Save txt, html, and pdf files generated from the above tasks in GCP Cloud Storage with daily or hourly date-time prefixes.
- Use LangChain and an embedding model (e.g. OpenAI’s embedding API) to create embeddings for the documents.
4. Document Embeddings with LangChain and OpenAI: Embed and index documents using LangChain’s RecordManager in incremental mode. LangChain’s RecordManager offers two modes for indexing: incremental and full. Both modes provide automated cleanup:
- If the content of the source document or derived documents has changed, both incremental and full modes will clean up (delete) previous versions of the content.
- If the source document has been deleted (meaning it is not included in the documents currently being indexed), the full cleanup mode will delete it from the vector store correctly, but the incremental mode will not.
- Redis Vector Store: Store the embeddings for future search and retrieval tasks.
- Web App hosted in Cloud Run:
- Use LangChain and Streamlit to build a question-answering web app based on the indexed documents.
- Utilize a document retriever and a LLM model (e.g ChatGPT) to generate answers.
Section 2: Local Testing for PoC
To test this document indexing and Q&A approach locally, we’ll set up a Docker environment with Redis and PostgreSQL as the core storage components. This PoC will include scripts to simulate data ingestion, embedding, and retrieval functionalities. Additionally, we’ll create sample documents in a dedicated folder to act as the initial content, simulate data ingestion from Confluence with a fetch_confluence_pages.py script, index documents with document_indexer.py, and implement a basic Q&A functionality using main.py.
Project Structure
Here’s how the project structure will look:
project-folder/
├── docker-compose.yml
├── documents/
│ ├── doc1.html
│ ├── doc2.txt
│ └── doc3.txt
├── fetch_confluence_pages.py
├── document_indexer.py
└── main.py
Step 1: Create Sample Documents
Create a local directory named documents and place some sample text files there.
mkdir documents
cd documents
echo "This is a sample document about AI." > doc2.txt
echo "Another document related to machine learning." > doc3.txt
cat <<EOF > doc1.html
<!DOCTYPE html>
<html>
<head>
<title>Understanding LLM Parameters</title>
</head>
<body>
<h1>Understanding LLM Parameters</h1>
<p>
LLM parameters, often referred to as weights, are crucial for the model's ability to understand and generate human-like text.
They are fine-tuned during training to minimize errors in predictions, ensuring that the model responds accurately to a wide range of inputs.
</p>
</body>
</html>
EOF
Step 2: Set Up Docker Compose for Redis and PostgreSQL
To create the local environment, use Docker Compose to set up Redis and PostgreSQL. This setup will help you test the document indexing pipeline with vector embedding storage and metadata storage for RecordManager.
Create a docker-compose.yml file:
version: '3.8'
services:
postgres:
image: postgres:15
container_name: local-postgres
restart: always
environment:
POSTGRES_USER: myuser
POSTGRES_PASSWORD: mypassword
POSTGRES_DB: mydb
ports:
- "5432:5432"
volumes:
- postgres_data:/var/lib/postgresql/data
networks:
- app-network
redis:
image: redis/redis-stack:latest
container_name: local-redis
restart: always
ports:
- "6379:6379"
volumes:
- redis_data:/data
networks:
- app-network
volumes:
postgres_data:
driver: local
redis_data:
driver: local
networks:
app-network:
driver: bridge
then run:
docker-compose up -d
Step 3: Install Required Libraries
langchain==0.3.4
langchain-community==0.3.3
langchain-core==0.3.13
langchain-openai==0.2.3
langchain-redis==0.1.1
psycopg2-binary
pypdf
dotenv
unstructured
#rank_bm25
Step 4: fetch_confluence_pages.py – Data ingestion from Confluence (optional)
This script fetches documents from Confluence and stores them in the documents folder. It connects to Confluence’s REST API to retrieve pages created within the last 24 hours.
fetch_confluence_pages.py
import os
from dotenv import load_dotenv
import logging
import requests
from google.cloud import storage
from datetime import datetime, timedelta
from typing import List, Dict, Optional, Tuple
import base64
from urllib.parse import quote
from pathlib import Path
load_dotenv()
# Configure logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
CONFLUENCE_DOMAIN = os.getenv('CONFLUENCE_DOMAIN')
CONFLUENCE_API_USER = os.getenv('CONFLUENCE_API_USER')
CONFLUENCE_EMAIL = os.getenv('CONFLUENCE_EMAIL')
CONFLUENCE_API_TOKEN = os.getenv('CONFLUENCE_API_TOKEN')
GCP_BUCKET_NAME = os.getenv('GCP_BUCKET_NAME')
google_credentials_path = os.getenv('GOOGLE_APPLICATION_CREDENTIALS')
LOCAL_DOCUMENTS_DIR = 'documents'
CONFLUENCE_BASE_URL = f"https://{CONFLUENCE_DOMAIN}"
def get_auth_headers():
"""Generate the authentication headers for Confluence API."""
auth_string = f"{CONFLUENCE_EMAIL}:{CONFLUENCE_API_TOKEN}"
auth_bytes = auth_string.encode('ascii')
base64_auth = base64.b64encode(auth_bytes).decode('ascii')
return {
'Authorization': f'Basic {base64_auth}',
'Accept': 'application/json'
}
def validate_environment() -> bool:
"""Validate that all required environment variables are set."""
required_vars = [
'CONFLUENCE_DOMAIN',
'CONFLUENCE_EMAIL',
'CONFLUENCE_API_TOKEN',
'GCP_BUCKET_NAME',
'GOOGLE_APPLICATION_CREDENTIALS'
]
missing_vars = [var for var in required_vars if not os.getenv(var)]
if missing_vars:
logger.error(
f"Missing required environment variables: {', '.join(missing_vars)}")
return False
return True
def get_storage_client() -> Optional[storage.Client]:
"""Initialize and return GCP storage client."""
try:
return storage.Client()
except Exception as e:
logger.error(f"Failed to initialize GCP storage client: {str(e)}")
return None
def get_date_folders() -> Tuple[str, str]:
"""
Get the folder paths for storing files.
Returns both the GCS path and the date string for the CQL query.
"""
yesterday = datetime.now() - timedelta(days=1)
folder_date = yesterday.strftime('%Y-%m-%d')
gcs_folder = f"documents/{folder_date}"
return gcs_folder, folder_date
def fetch_new_confluence_pages() -> None:
"""Fetch new Confluence pages and save them to GCP Storage."""
if not validate_environment():
return
storage_client = get_storage_client()
if not storage_client:
return
try:
# Get folder paths
gcs_folder, query_date = get_date_folders()
# Get bucket and list existing pages
bucket = storage_client.bucket(GCP_BUCKET_NAME)
logger.info(f"Storing files in folder: {gcs_folder}")
# Construct CQL query
cql_query = quote(f"type=page and created >= '{query_date}'")
url = f"{CONFLUENCE_BASE_URL}/wiki/rest/api/content/search?cql={cql_query}&expand=body.storage"
logger.info(f"Requesting URL: {url}")
# Using headers for authentication
headers = get_auth_headers()
response = requests.get(url, headers=headers, timeout=30)
# Debug logging
logger.info(f"API Response Status: {response.status_code}")
if response.status_code != 200:
logger.error(f"API Response: {response.text}")
response.raise_for_status()
pages = response.json().get('results', [])
logger.info(f"Found {len(pages)} pages to process")
processed_count = 0
skipped_count = 0
for page in pages:
try:
was_processed = process_single_page(
page, bucket, headers, gcs_folder)
if was_processed:
processed_count += 1
else:
skipped_count += 1
except Exception as e:
logger.error(
f"Failed to process page {page.get('id', 'unknown')}: {str(e)}")
continue
logger.info(
f"Processing complete. Processed: {processed_count}, Skipped: {skipped_count}")
except requests.exceptions.RequestException as e:
logger.error(f"Failed to fetch Confluence pages: {str(e)}")
logger.error(
f"Response content: {getattr(e.response, 'text', 'No response content')}")
except Exception as e:
logger.error(f"Unexpected error: {str(e)}")
def process_single_page(page: Dict, bucket: storage.Bucket, headers: Dict,
folder_path: str) -> bool:
"""
Process a single Confluence page and save it to GCP Storage.
Returns True if page was processed, False if skipped.
"""
page_id = page['id']
title = page['title']
# Sanitize filename
safe_title = "".join(c for c in title if c.isalnum()
or c in (' ', '-', '_')).rstrip()
filename = f"{safe_title}.html"
logger.info(f"Processing page: {safe_title} (ID: {page_id})")
content_url = f"{CONFLUENCE_BASE_URL}/wiki/rest/api/content/{page_id}?expand=body.storage"
content_response = requests.get(content_url, headers=headers, timeout=30)
content_response.raise_for_status()
content = content_response.json()['body']['storage']['value']
# Save content to GCP Storage with metadata
blob = bucket.blob(f"{folder_path}/{filename}")
# Add metadata
metadata = {
'confluence_id': page_id,
'original_title': title,
'backup_date': datetime.now().isoformat(),
'content_type': 'confluence_page'
}
blob.metadata = metadata
blob.upload_from_string(content, content_type='text/html')
logger.info(f"Successfully saved page: {safe_title}")
# Save content to local documents folder for local run only
local_folder = Path(LOCAL_DOCUMENTS_DIR)
local_file_path = local_folder / filename
with open(local_file_path, 'w', encoding='utf-8') as file:
file.write(content)
logger.info(f"Successfully saved page to local folder: {safe_title}")
return True
if __name__ == "__main__":
fetch_new_confluence_pages()
Step 5: document_indexer.py – Index Documents Using Embeddings
This script will load all documents in the documents folder, split each document as needed, generate embeddings for each segment, and store them in the Redis as vector database.
Langchain RecordManager checks the metadata in the Postgres db upsertion_record table before indexing, tracking each document’s last modification date. When run for the first time, it indexes all documents and updates the metadata. For subsequent runs, it only indexes new or modified documents.
document_indexer.py
import os
import logging
from typing import List, Dict, Optional
from dataclasses import dataclass
from pathlib import Path
from datetime import datetime
from dotenv import load_dotenv
from langchain.indexes import SQLRecordManager, index
from langchain.schema import Document
from langchain_community.document_loaders import (
TextLoader,
UnstructuredHTMLLoader,
PyPDFLoader
)
from langchain_openai import OpenAIEmbeddings
from langchain.vectorstores import Redis
from langchain_text_splitters import RecursiveCharacterTextSplitter
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
handlers=[
logging.FileHandler('document_indexing.log'),
logging.StreamHandler()
]
)
logger = logging.getLogger(__name__)
@dataclass
class IndexingConfig:
"""Configuration for document indexing system."""
documents_dir: str
redis_host: str
redis_port: int
postgres_host: str
postgres_port: int
postgres_db: str
postgres_user: str
postgres_password: str
openai_api_key: str
chunk_size: int = 1000
chunk_overlap: int = 200
@property
def redis_url(self) -> str:
return f"redis://{self.redis_host}:{self.redis_port}"
@property
def postgres_url(self) -> str:
return f"postgresql+psycopg2://{self.postgres_user}:{self.postgres_password}@{self.postgres_host}:{self.postgres_port}/{self.postgres_db}"
@classmethod
def from_env(cls) -> 'IndexingConfig':
"""Create configuration from environment variables."""
load_dotenv()
return cls(
documents_dir=os.getenv('DOCUMENTS_DIR', 'documents'),
redis_host=os.getenv('REDIS_HOST', 'localhost'),
redis_port=int(os.getenv('REDIS_PORT', '6379')),
postgres_host=os.getenv('POSTGRES_HOST', 'localhost'),
postgres_port=int(os.getenv('POSTGRES_PORT', '5432')),
postgres_db=os.getenv('POSTGRES_DB', 'mydb'),
postgres_user=os.getenv('POSTGRES_USER'),
postgres_password=os.getenv('POSTGRES_PASSWORD'),
openai_api_key=os.getenv('OPENAI_API_KEY'),
chunk_size=int(os.getenv('CHUNK_SIZE', '1000')),
chunk_overlap=int(os.getenv('CHUNK_OVERLAP', '200'))
)
class DocumentProcessor:
"""Handles document loading and processing."""
SUPPORTED_EXTENSIONS = {
'.txt': TextLoader,
'.html': UnstructuredHTMLLoader,
'.htm': UnstructuredHTMLLoader,
'.pdf': PyPDFLoader
}
def __init__(self, config: IndexingConfig):
self.config = config
self.splitter = RecursiveCharacterTextSplitter(
separators=["\n\n", "\n", " ", ""],
chunk_size=config.chunk_size,
chunk_overlap=config.chunk_overlap,
length_function=len
)
def load_documents(self) -> List[Document]:
"""Load documents from the specified directory."""
documents = []
directory_path = Path(self.config.documents_dir)
if not directory_path.exists():
logger.error(f"Directory not found: {directory_path}")
raise FileNotFoundError(f"Directory not found: {directory_path}")
for file_path in directory_path.iterdir():
if not file_path.is_file():
continue
ext = file_path.suffix.lower()
loader_class = self.SUPPORTED_EXTENSIONS.get(ext)
if not loader_class:
logger.warning(
f"Skipping unsupported file format: {file_path.name}")
continue
try:
logger.info(f"Loading document: {file_path.name}")
loader = loader_class(str(file_path))
docs = loader.load()
# Add metadata
for doc in docs:
doc.metadata = {"source": file_path.name}
documents.extend(docs)
except Exception as e:
logger.error(f"Failed to load {file_path.name}: {str(e)}")
continue
logger.info(f"Successfully loaded {len(documents)} documents")
return documents
class DocumentIndexer:
"""Handles document indexing and embedding storage."""
def __init__(self, config: IndexingConfig):
self.config = config
self.processor = DocumentProcessor(config)
def initialize_components(self):
"""Initialize embeddings, vectorstore, and record manager."""
try:
self.embeddings = OpenAIEmbeddings(
openai_api_key=self.config.openai_api_key)
self.vectorstore = Redis(
redis_url=self.config.redis_url,
index_name='document_embeddings',
embedding=self.embeddings
)
self.record_manager = SQLRecordManager(
namespace="redis/document_embeddings",
db_url=self.config.postgres_url
)
# Create the schema if it doesn't exist
self.record_manager.create_schema()
logger.info("Successfully initialized all components")
except Exception as e:
logger.error(f"Failed to initialize components: {str(e)}")
raise
def process_and_index_documents(self):
"""Process and index documents with error handling and logging."""
try:
documents = self.processor.load_documents()
if not documents:
logger.warning("No documents were loaded")
return
logger.info("Splitting documents into chunks...")
chunks = self.processor.splitter.split_documents(documents)
logger.info(
f"Created {len(chunks)} chunks from {len(documents)} documents")
# Index documents
logger.info("Indexing documents...")
result = index(
chunks,
self.record_manager,
self.vectorstore,
cleanup="incremental",
source_id_key="source",
)
logger.info("Indexing completed successfully")
logger.info(f"Indexing results: {result}")
return result
except Exception as e:
logger.error(
f"Error during document processing and indexing: {str(e)}")
raise
def main():
"""Main function to run the document indexing process."""
try:
config = IndexingConfig.from_env()
indexer = DocumentIndexer(config)
indexer.initialize_components()
# Process and index documents
result = indexer.process_and_index_documents()
logger.info("Document indexing process completed successfully")
return result
except Exception as e:
logger.error(f"Document indexing process failed: {str(e)}")
raise
if __name__ == "__main__":
main()
Expected Output on First and Subsequent Runs
- First Run: Since there’s no existing metadata, all documents will be indexed.
{'num_added': 4, 'num_updated': 0, 'num_skipped': 0, 'num_deleted': 0}
- Subsequent Runs: If there are no new or modified documents, the output will indicate that all documents were skipped.
{'num_added': 0, 'num_updated': 0, 'num_skipped': 4, 'num_deleted': 0}
Step 6:main.py – Set Up Q&A Retrieval with LangChain
The script provides a simple interface to query documents in Redis using LangChain to respond to user questions.
main.py
import os
import logging
from typing import List, Optional, Dict
from dataclasses import dataclass
from langchain.vectorstores import Redis
from langchain.schema import Document
from dotenv import load_dotenv
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import OpenAIEmbeddings
from langchain_openai.chat_models import ChatOpenAI
from langchain_core.prompts import PromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.retrievers import BaseRetriever
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
@dataclass
class Config:
"""Configuration class for the application."""
redis_host: str
redis_port: int = 6379
redis_index_name: str = "document_embeddings"
retriever_k: int = 2
max_retries: int = 3
temperature: float = 0
@property
def redis_url(self) -> str:
return f"redis://{self.redis_host}:{self.redis_port}"
@classmethod
def from_env(cls) -> 'Config':
"""Create configuration from environment variables."""
load_dotenv()
return cls(
redis_host=os.getenv('REDIS_HOST', 'localhost'),
redis_port=int(os.getenv('REDIS_PORT', '6379')),
redis_index_name=os.getenv(
'REDIS_INDEX_NAME', 'document_embeddings'),
retriever_k=int(os.getenv('RETRIEVER_K', '2')),
max_retries=int(os.getenv('MAX_RETRIES', '3')),
temperature=float(os.getenv('TEMPERATURE', 0))
)
class QASystem:
"""Main QA system class handling document retrieval and question answering."""
DEFAULT_PROMPT_TEMPLATE = """
You are an assistant for question-answering tasks. Use the following pieces of retrieved context
to answer the question. If you don't know the answer, just say that you don't know.
Use three sentences maximum and keep the answer concise.
Question: {question}
Context: {context}
Answer:
"""
def __init__(self, config: Config):
self.config = config
self.embedding = OpenAIEmbeddings()
self.initialize_components()
def initialize_components(self):
"""Initialize Redis vectorstore, retriever, and QA chain."""
try:
self.redis_vectorstore = Redis(
redis_url=self.config.redis_url,
index_name=self.config.redis_index_name,
embedding=self.embedding
)
logger.info(
f"Successfully connected to Redis at {self.config.redis_url}")
self.retriever = self.initialize_retriever()
self.qa_chain = self.create_qa_chain()
except Exception as e:
logger.error(f"Failed to initialize components: {str(e)}")
raise
def initialize_retriever(self) -> BaseRetriever:
"""Initialize the Redis retriever."""
return self.redis_vectorstore.as_retriever(
search_kwargs={"k": self.config.retriever_k}
)
@staticmethod
def format_docs(docs: List[Document]) -> str:
"""Format the retrieved documents into a single string."""
return "\n\n".join(doc.page_content for doc in docs)
def create_qa_chain(self):
"""Create the question-answering chain."""
prompt = PromptTemplate.from_template(self.DEFAULT_PROMPT_TEMPLATE)
llm = ChatOpenAI(temperature=self.config.temperature)
return (
{
"context": self.retriever | self.format_docs,
"question": RunnablePassthrough(),
}
| prompt
| llm
| StrOutputParser()
)
def answer_question(self, question: str, max_retries: Optional[int] = None) -> str:
"""
Answer a question using the QA chain with retry mechanism.
Args:
question: The question to answer
max_retries: Maximum number of retry attempts
Returns:
str: The answer to the question
"""
max_retries = max_retries or self.config.max_retries
retry_count = 0
while retry_count < max_retries:
try:
logger.info(f"Processing question: {question}")
answer = self.qa_chain.invoke(question)
logger.info("Successfully generated answer")
return answer
except Exception as e:
retry_count += 1
if retry_count < max_retries:
logger.warning(
f"Attempt {retry_count} failed: {str(e)}. Retrying...")
else:
logger.error(
f"Failed to answer question after {max_retries} attempts")
raise RuntimeError(f"Failed to answer question: {str(e)}")
def get_document_count(self) -> int:
"""Get the number of documents in the Redis vectorstore."""
try:
# Using a broad search to get document count
docs = self.redis_vectorstore.similarity_search("", k=1)
return len(docs)
except Exception as e:
logger.error(f"Failed to get document count: {str(e)}")
return 0
def interactive_qa(self):
"""
Run an interactive Q&A session where users can input questions
until they type 'exit'.
"""
print("\nWelcome to the Interactive Q&A System!")
print("Type 'exit' to quit the program")
print("Type 'help' for available commands")
print("-" * 50)
while True:
try:
# Get user input
question = input("\nEnter your question: ").strip()
# Check for exit command
if question.lower() == 'exit':
print("\nThank you for using the Q&A system. Goodbye!")
break
# Check for help command
elif question.lower() == 'help':
print("\nAvailable commands:")
print("- 'exit': Quit the program")
print("- 'help': Show this help message")
print("- 'count': Show number of documents in the system")
continue
# Check for count command
elif question.lower() == 'count':
doc_count = self.get_document_count()
print(f"\nNumber of documents in the system: {doc_count}")
continue
# Skip empty questions
if not question:
print("Please enter a valid question.")
continue
# Get and display answer
print("\nProcessing your question...")
answer = self.answer_question(question)
print("\nAnswer:", answer)
print("-" * 50)
except KeyboardInterrupt:
print("\n\nProgram interrupted by user. Goodbye!")
break
except Exception as e:
logger.error(f"Error processing question: {str(e)}")
print(
"\nSorry, there was an error processing your question. Please try again.")
def main():
"""Main function to run the QA system."""
try:
config = Config.from_env()
qa_system = QASystem(config)
doc_count = qa_system.get_document_count()
logger.info(f"System initialized with {doc_count} documents in Redis")
qa_system.interactive_qa()
except Exception as e:
logger.error(f"Application error: {str(e)}")
raise
if __name__ == "__main__":
main()
After running the optional fetch_confluence_pages.py (to pull recent documents from Confluence) and executing document_indexer.py to embed and index documents in Redis, you can start the interactive Q&A interface by running main.py.
Once you run main.py, you will be prompted to enter a question. The system uses embeddings to find relevant information from the indexed documents and then leverages LangChain and ChatGPT to generate a response.
Summary
That’s about it! From this article, you should have a basic understanding of building a solution for continuous document indexing. For the steps outlined, you can implement a similar pipeline, adapt it to your needs, and scale it to handle large volumes of documents and user queries.
文章来源: https://hackernoon.com/behind-every-question-answer-ai-is-a-data-pipeline-built-for-scale-heres-how-to-build-your-own?source=rss
如有侵权请联系:admin#unsafe.sh
如有侵权请联系:admin#unsafe.sh