Welcome to this comprehensive guide on constructing an AI-enhanced Chat Application. We will create a dynamic chat interface capable of delving into PDF documents to extract and provide summaries, key facts, and answers to your queries. By the end of this tutorial, you’ll have a powerful tool at your disposal, transforming the way you interact with and utilize the information contained within PDFs.
This tutorial will demonstrate how to -
Note that this tutorial is designed to work with the latest Python SDK version (4.2.0+) for Couchbase. It will not work with the older Python SDK versions.
Vector Search in Couchbase is only supported at Couchbase Version 7.6+.
git clone https://github.com/couchbase-examples/rag-gemini-demo.git
Any dependencies should be installed through pip
, the default package manager for Python. You may use virtual environment as well.
python -m pip install -r requirements.txt
To know more about connecting to your Capella cluster, please follow the instructions.
Specifically, you need to do the following:
pdf-chat
. We will use the _default
scope and _default
collection of this bucket.We need to create the Search Index on the Full Text Service in Couchbase. For this demo, you can import the following index using the instructions.
You may also create a vector index using Search UI on both Couchbase Capella and Couchbase Self Managed Server.
Here, we are creating the index pdf_search
on the documents. The Vector field is set to embedding
with 768 dimensions and the text field set to text
. We are also indexing and storing all the fields under metadata
in the document as a dynamic mapping to account for varying document structures. The similarity metric is set to dot_product
. If there is a change in these parameters, please adapt the index accordingly.
{
"name": "pdf_search",
"type": "fulltext-index",
"params": {
"doc_config": {
"docid_prefix_delim": "",
"docid_regexp": "",
"mode": "scope.collection.type_field",
"type_field": "type"
},
"mapping": {
"default_analyzer": "standard",
"default_datetime_parser": "dateTimeOptional",
"default_field": "_all",
"default_mapping": {
"dynamic": true,
"enabled": false
},
"default_type": "_default",
"docvalues_dynamic": false,
"index_dynamic": true,
"store_dynamic": false,
"type_field": "_type",
"types": {
"_default._default": {
"dynamic": true,
"enabled": true,
"properties": {
"embedding": {
"enabled": true,
"dynamic": false,
"fields": [
{
"dims": 768,
"index": true,
"name": "embedding",
"similarity": "dot_product",
"type": "vector",
"vector_index_optimized_for": "recall"
}
]
},
"metadata": {
"dynamic": true,
"enabled": true
},
"text": {
"enabled": true,
"dynamic": false,
"fields": [
{
"index": true,
"name": "text",
"store": true,
"type": "text"
}
]
}
}
}
}
},
"store": {
"indexType": "scorch",
"segmentVersion": 16
}
},
"sourceType": "gocbcore",
"sourceName": "pdf-chat",
"sourceParams": {},
"planParams": {
"maxPartitionsPerPIndex": 64,
"indexPartitions": 16,
"numReplicas": 0
}
}
Copy the secrets.example.toml
file in .streamlit
folder and rename it to secrets.toml and replace the placeholders with the actual values for your environment. All configuration for communication with the database is read from the environment variables.
GOOGLE_API_KEY = "<google_ai_api_key>"
DB_CONN_STR = "<connection_string_for_couchbase_cluster>"
DB_USERNAME = "<username_for_couchbase_cluster>"
DB_PASSWORD = "<password_for_couchbase_cluster>"
DB_BUCKET = "<name_of_bucket_to_store_documents>"
DB_SCOPE = "<name_of_scope_to_store_documents>"
DB_COLLECTION = "<name_of_collection_to_store_documents>"
INDEX_NAME = "<name_of_fts_index_with_vector_support>"
LOGIN_PASSWORD = "<password to access the streamlit app>"
Google AI API Key is required for usage in generating embedding and querying LLM.
The connection string expects the
couchbases://
orcouchbase://
part.
For this tutorial,
DB_BUCKET = pdf-chat
,DB_SCOPE = _default
,DB_COLLECTION = _default
andINDEX_NAME = pdf_search
.
Login_Password of Streamlit app is a basic password to access the app. You can set the password here and while using the app, password will be required to access the app.
After starting Couchbase server, adding vector index and installing dependencies. Our Application is ready to run.
In the projects root directory, run the following command
streamlit run chat_with_pdf.py
The application will run on your local machine at http://localhost:8501.
The page starts with a text box to enter your login password. This is the same password defined earlier in the Setup Environment Config section. After submitting the password, the main application page will open.
On the left sidebar, you'll find an option to upload a PDF document you want to use with this PDF Chat App. Depending on the size of the PDF, the upload process may take some time.
In the main area, there's a chat screen where you can ask questions about the uploaded PDF document. You will receive two responses: one with context from the PDF (Couchbase Logo - ) , and one without the PDF context (Bot Logo - 🤖). This demonstrates how the Retrieval Augmented Generation (RAG) model enhances the answers provided by the language model using the PDF content.
The PDF Chat application leverages two powerful concepts: Retrieval-Augmented Generation (RAG) and Vector Search. Together, these techniques enable efficient and context-aware interactions with PDF documents.
RAG is like having two helpers:
Here's how RAG works:
This enhances the context from PDF and LLM is able to give relevant results from the PDF rather than giving generalized results.
Couchbase is a NoSQL database that provides a powerful Vector Search capability. It allows you to store and search through high-dimensional vector representations (embeddings) of textual data, such as PDF content.
The PDF Chat app uses LangChain to convert the text from the PDF documents into embeddings. These embeddings are then stored in a Couchbase bucket, along with the corresponding text.
When a user asks a question or provides a prompt:
LangChain is a powerful library that simplifies the process of building applications with large language models (LLMs) and vector stores like Couchbase.
In the PDF Chat app, LangChain is used for several tasks:
By combining Vector Search with Couchbase, RAG, and LangChain; the PDF Chat app can efficiently ingest PDF documents, convert their content into searchable embeddings, retrieve relevant information based on user queries and conversation context, and generate context-aware and informative responses using large language models. This approach provides users with a powerful and intuitive way to explore and interact with large PDF files.
To begin this tutorial, clone the repo and open it up in the IDE of your choice. Now you can learn how to create the PDF Chat App. The whole code is written in chat_with_pdf.py
file.
The fundamental workflow of the application is as follows: The user initiates the process from the Main Page's sidebar by uploading a PDF. This action triggers the save_to_vector_store
function, which subsequently uploads the PDF into the Couchbase vector store. Following this, the user can now chat with the LLM.
On the Chat Area, the user can pose questions. These inquiries are processed by the Chat API, which consults the LLM for responses, aided by the context provided by RAG. The assistant then delivers the answer, and the user has the option to ask additional questions.
The first step will be connecting to Couchbase. Couchbase Vector Search is required for PDF Upload as well as during chat (For Retrieval). We will use the Couchbase Python SDK to connect to the Couchbase cluster. The connection is established in the connect_to_couchbase
function.
The connection string and credentials are read from the environment variables. We perform some basic required checks for the environment variable not being set in the secrets.toml
, and then proceed to connect to the Couchbase cluster. We connect to the cluster using connect method.
def connect_to_couchbase(connection_string, db_username, db_password):
"""Connect to Couchbase"""
from couchbase.cluster import Cluster
from couchbase.auth import PasswordAuthenticator
from couchbase.options import ClusterOptions
from datetime import timedelta
auth = PasswordAuthenticator(db_username, db_password)
options = ClusterOptions(auth)
connect_string = connection_string
cluster = Cluster(connect_string, options)
# Wait until the cluster is ready for use.
cluster.wait_until_ready(timedelta(seconds=5))
return cluster
We will now initialize Google Generative AI Embeddings which will be used by CouchbaseVectorStore for converting the split docs defined above to vectors (embeddings).
We will also initialize Couchbase vector store with Couchbase bucket info. Firstly we will connect to Couchbase cluster using connect_to_couchbase
method.
We will define the bucket, scope, collection and index names from Environment Variables.
# Use Google Generative AI Embeddings
embedding = GoogleGenerativeAIEmbeddings(
model="models/text-embedding-004",
)
# Connect to Couchbase Vector Store
cluster = connect_to_couchbase(DB_CONN_STR, DB_USERNAME, DB_PASSWORD)
vector_store = get_vector_store(
cluster,
DB_BUCKET,
DB_SCOPE,
DB_COLLECTION,
embedding,
INDEX_NAME,
)
We are using get_vector_store
method which initializes LangChain's CouchbaseVectorStore
def get_vector_store(
_cluster,
db_bucket,
db_scope,
db_collection,
_embedding,
index_name,
):
"""Return the Couchbase vector store"""
vector_store = CouchbaseVectorStore(
cluster=_cluster,
bucket_name=db_bucket,
scope_name=db_scope,
collection_name=db_collection,
embedding=_embedding,
index_name=index_name,
)
return vector_store
save_to_vector_store
function takes care of uploading the PDF file in vector format to Couchbase Database using CouchbaseVectorStore in LangChain. It splits text into small chunks, generate embeddings for those chunks, and ingest the chunks and their embeddings into a Couchbase vector store. Let's go step by step on how it does.
This part of code creates a file uploader on sidebar using Streamlit library. After PDF is uploaded, save_to_vector_store
function is called to further process the PDF.
with st.form("upload pdf"):
uploaded_file = st.file_uploader(
"Choose a PDF.",
help="The document will be deleted after one hour of inactivity (TTL).",
type="pdf",
)
submitted = st.form_submit_button("Upload")
if submitted:
# store the PDF in the vector store after chunking
save_to_vector_store(uploaded_file, vector_store)
This function ensures that the uploaded PDF file is properly handled, loaded, and prepared for storage or processing in the vector store. It first checks if file was actually uploaded. Then the uploaded file is saved to a temporary file in binary
format.
From the temporary file, PDF is loaded in PyPDFLoader from the LangChain library which loads the PDF into LangChain Document Format
def save_to_vector_store(uploaded_file, vector_store):
"""Chunk the PDF & store it in Couchbase Vector Store"""
if uploaded_file is not None:
temp_dir = tempfile.TemporaryDirectory()
temp_file_path = os.path.join(temp_dir.name, uploaded_file.name)
with open(temp_file_path, "wb") as f:
f.write(uploaded_file.getvalue())
loader = PyPDFLoader(temp_file_path)
docs = loader.load()
We will utilize the vector store created at Initialize Embeddings and Couchbase Vector Store. In this we will add the documents using add_documents method of Couchbase vector store. This method will utilize the Google's generative AI embeddings to create embeddings(vectors) from text and add it to Couchbase documents in the specified collection.
vector_store.add_documents(docs)
After uploading the PDF into Couchbase, we are now ready to utilize the power of Couchbase Vector Search, RAG and LLM to get context based answers to our questions. When the user asks a question. The assistant (LLM) is called here with RAG context, the response from the assistant is sent back to the user.
We will now utilize the power of LangChain Chains using the LangChain Expression Language (LCEL). LCEL makes it easy to build complex chains from basic components, and supports out of the box functionality such as streaming, parallelism, and logging.
LCEL is a domain-specific language that provides several key advantages when working with LangChain:
In summary, LCEL streamlines the process of building sophisticated natural language processing applications by providing a composable, readable, reusable, extensible, type-safe, and abstracted way to define and orchestrate LangChain components into complex workflows.
We will be using LCEL chains in next few sections and will see how LCEL optimizes our whole workflow.
We also create the retriever of the couchbase vector store. This retriever will be used to retrieve the previously added documents which are similar to current query.
retriever = vector_store.as_retriever()
A prompt for a language model is a set of instructions or input provided by a user to guide the model's response, helping it understand the context and generate relevant and coherent language-based output, such as answering questions, completing sentences, or engaging in a conversation. We will use a template and create a prompt chain using ChatPromptTemplate Class of LangChain
template = """You are a helpful bot. If you cannot answer based on the context provided, respond with a generic answer. Answer the question as truthfully as possible using the context below:
{context}
Question: {question}"""
prompt = ChatPromptTemplate.from_template(template)
Large Language Models (LLMs) are a core component of LangChain. LangChain does not serve its own LLMs, but rather provides a standard interface for interacting with many LLMs. To be specific, this interface is one that takes as input a string and returns a string. We will use Google AI LLM Model with gemini model. We can also set other parameters like model, API_KEY, temperature to be used for this model.
llm = GoogleGenerativeAI(
temperature=0.1,
model="models/gemini-1.5-pro",
)
We can combine these different modules to a single chain which will run synchronously. The pipe operator (|) is used to connect these components, forming a chain of operations.
The input dictionary with the "context" and "question" keys is passed through each component. The retriever component retrieves relevant information from the vector store, which is then combined with the question to generate a prompt. This prompt is passed to the language model (llm), which generates a response based on both the question and the retrieved context. Finally, the output is parsed into a string by StrOutputParser().
# RAG chain
chain = (
{"context": retriever, "question": RunnablePassthrough()}
| prompt
| llm
| StrOutputParser()
)
We will repeat the same process as above however this will not have the context from the vector store. Basically we will directly call LLM from the user question. Basically, every step is same just that we will not use retriever.
template_without_rag = """You are a helpful bot. Answer the question as truthfully as possible.
Question: {question}"""
prompt_without_rag = ChatPromptTemplate.from_template(template_without_rag)
llm_without_rag = GoogleGenerativeAI(
temperature=0,
model="models/gemini-1.5-pro",
)
chain_without_rag = (
{"question": RunnablePassthrough()}
| prompt_without_rag
| llm_without_rag
| StrOutputParser()
)
This section creates an interactive chat interface where users can ask questions based on the uploaded PDF. The key steps are:
This setup allows users to have a conversational experience, asking questions related to the uploaded PDF, with responses generated by the RAG chain and streamed in real-time. Both the user's questions and the assistant's responses are displayed in the chat interface, along with their respective roles and avatars.
# React to user input
if question := st.chat_input("Ask a question based on the PDF"):
# Display user message in chat message container
st.chat_message("user").markdown(question)
# Add user message to chat history
st.session_state.messages.append(
{"role": "user", "content": question, "avatar": "👤"}
)
# Add placeholder for streaming the response
with st.chat_message("assistant", avatar=couchbase_logo):
message_placeholder = st.empty()
# stream the response from the RAG
rag_response = ""
for chunk in chain.stream(question):
rag_response += chunk
message_placeholder.markdown(rag_response + "▌")
message_placeholder.markdown(rag_response)
st.session_state.messages.append(
{
"role": "assistant",
"content": rag_response,
"avatar": couchbase_logo,
}
)
Similar to last section, we will get answer from LLM of the user question. Answers from here are also shown in the UI to showcase difference on how using RAG gives better and more context enabled results.
# stream the response from the pure LLM
# Add placeholder for streaming the response
with st.chat_message("ai", avatar="🤖"):
message_placeholder_pure_llm = st.empty()
pure_llm_response = ""
for chunk in chain_without_rag.stream(question):
pure_llm_response += chunk
message_placeholder_pure_llm.markdown(pure_llm_response + "▌")
message_placeholder_pure_llm.markdown(pure_llm_response)
st.session_state.messages.append(
{
"role": "assistant",
"content": pure_llm_response,
"avatar": "🤖",
}
)