Lesson 3: 03 - Embeddings into Vector Database with FAISS

In this lesson, we will focus on this part of our global plan:

With the help of LangChain, we don't need to build the embeddings manually and call the embed_documents() function as we did in the last lesson.

We can create a vector database object and store the embeddings directly from the documents:

from langchain.document_loaders.csv_loader import CSVLoader
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores import FAISS 
 
loader = CSVLoader(file_path="faq.csv")
documents = loader.load()
 
embeddings_model = OpenAIEmbeddings(openai_api_key='sk-...')
 
vectorstore = FAISS.from_documents(documents, embeddings_model) 

Wait, what's that FAISS?

OpenAI provides the embedding mechanism but not the space to store those embeddings. So, we need to choose from other additional tools. Similarly to embeddings, there are plenty of options for storing vectors:

• FAISS by Facebook (we will use it in this tutorial)
• Pinecone
• Chroma
• Weaviate
• ... many more

Some of those are specific vector databases, others are more general database systems that can store vectors. You can check the most popular vector databases in the article LangChain State of AI 2023.

In our case, we will use FAISS. You must install it in your system with pip install faiss-cpu. There's also a faiss-gpu version if you have a powerful enough GPU and want to utilize it.

As a result, we now have a vectorstore object which allows us to perform the similarity search in a vector database:

results = vectorstore.similarity_search('Do you have a free plan?')
print(results)

# Output:
[
  Document(page_content='Question: What are your pricing plans?\nAnswer: What are your pricing plans?\n ...',
    metadata={'source': 'faq.csv', 'row': 0}),
  Document(page_content="Question: Do I get charged automatically for upcoming months/years?\nAnswer: We use Paddle ...",
     metadata={'source': 'faq.csv', 'row': 2}),
  Document(page_content='Question: Any discount coupons based on country? PPP?\nAnswer: Yes. All Purchasing Power Parity ...',
     metadata={'source': 'faq.csv', 'row': 5}),
  Document(page_content='Question: What am I getting as a Premium Member?\nAnswer: You will get access to:\n- 40+ video...',
     metadata={'source': 'faq.csv', 'row': 1})
]

As you can see, the similarity search returns the documents from the list above that were found to be similar to the text query. Again, the search is based on vector embeddings, not string texts.

But for the chatbot, we won't need to perform this similarity_search() manually. Again, LangChain will do it for us behind the scenes. We just need to provide the vectorstore variable, and LangChain will call the LLM (OpenAI GPT-3.5-Turbo), passing those filtered documents as so-called context.

To provide the answer to the chatbot query, the OpenAI GPT will use the internal information it has been trained on but also will use our documents as additional (or, rather, primary) context.

We will implement all of that in the next lesson. So, time to actually use the OpenAI API?