Enabling semantic search on user-specific data is a multi-step process that includes loading, transforming, embedding and storing data before it can be queried.

That graphic is from the team over at LangChain, whose goal is to provide a set of utilities to greatly simplify this process.

In this tutorial, we'll walk through each of these steps, using MongoDB Atlas as our Store. Specifically, we'll use the AT&T Wikipedia page as our data source. We'll then use libraries from LangChain to Load, Transform, Embed and Store:

Once the source is store is stored in MongoDB, we can retrieve the data that interests us:

• MongoDB Atlas Subscription (Free Tier is fine)
• Open AI API key

1. Get the code:

git clone https://github.com/wbleonard/atlas-langchain.git

2. Update params.py with your MongoDB connection string and Open AI API key.
3. Create a new Python environment

python3 -m venv env

4. Activate the new Python environment

source env/bin/activate

5. Install the requirements

pip3 install -r requirements.txt

6. Load, Transform, Embed and Store

python3 vectorize.py

7. Retrieve

python3 query.py -q "Who started AT&T?"

There's no lacking for sources of data: Slack, YouTube, Git, Excel, Reddit, Twitter, etc., and LangChain provides a growing list of integrations that includes this list and many more.

For this exercise, we're going to use the WebBaseLoader to load the Wikipedia page for AT&T.

from langchain.document_loaders import WebBaseLoader
loader = WebBaseLoader("https://en.wikipedia.org/wiki/AT%26T")
data = loader.load()

Now that we have a bunch of text loaded, it needs to be split into smaller chunks so we can tease out the relevant portion based on our search query. For this example we'll use the recommended RecursiveCharacterTextSplitter. As I have it configured, it attempts to split on paragraphs ("\n\n"), then sentences("(?<=\. )"), then words (" ") using a chunk size of 1000 characters. So if a paragraph doesn't fit into 1000 characters, it will truncate at the next word it can fit to keep the chunk size under 1000 chacters. You can tune the chunk_size to your liking. Smaller numbers will lead to more documents, and vice-versa.

from langchain.text_splitter import RecursiveCharacterTextSplitter
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0, separators=[
                                               "\n\n", "\n", "(?<=\. )", " "], length_function=len)
docs = text_splitter.split_documents(data)

Embedding is where you associate your text with an LLM to create a vector representation of that text. There are many options to choose from, such as OpenAI and Hugging Face, and LangChang provides a standard interface for interacting with all of them.

For this exercise we're going to use the popular OpenAI embedding. Before proceeding, you'll need an API key for the OpenAI platform, which you will set in params.py.

We're simply going to load the embedder in this step. The real power comes when we store the embeddings in Step 4.

from langchain.embeddings.openai import OpenAIEmbeddings
embeddings = OpenAIEmbeddings(openai_api_key=params.openai_api_key)

You'll need a vector database to store the embeddings, and lucky for you MongoDB fits that bill. Even luckier for you, the folks at LangChain have a MongoDB Atlas module that will do all the heavy lifting for you! Don't forget to add your MongoDB Atlas connection string to params.py.

from pymongo import MongoClient
from langchain.vectorstores import MongoDBAtlasVectorSearch

client = MongoClient(params.mongodb_conn_string)
collection = client[params.db_name][params.collection_name]

# Insert the documents in MongoDB Atlas with their embedding
docsearch = MongoDBAtlasVectorSearch.from_documents(
    docs, embeddings, collection=collection, index_name=index_name
)

You'll find the complete script in vectorize.py, which needs to be run once per data source (and you could easily modify the code to iterate over multiple data sources).

python3 vectorize.py

The final step before we can query the data is to create a search index on the stored embeddings.

In the Atlas console, create a Search Index using the JSON Editor named vsearch_index with the following definition:

{
  "mappings": {
    "dynamic": true,
    "fields": {
      "embedding": {
        "dimensions": 1536,
        "similarity": "cosine",
        "type": "knnVector"
      }
    }
  }
}

We could now run a search, using methods like similirity_search or max_marginal_relevance_search and that would return the relevant slice of data, which in our case would be an entire paragraph. However, we can continue to harness the power of the LLM to contextually compress the response so that it more directly tries to answer our question.

from pymongo import MongoClient
from langchain.vectorstores import MongoDBAtlasVectorSearch
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.llms import OpenAI
from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import LLMChainExtractor

client = MongoClient(params.mongodb_conn_string)
collection = client[params.db_name][params.collection_name]

vectorStore = MongoDBAtlasVectorSearch(
    collection, OpenAIEmbeddings(openai_api_key=params.openai_api_key), index_name=params.index_name
)

llm = OpenAI(openai_api_key=params.openai_api_key, temperature=0)
compressor = LLMChainExtractor.from_llm(llm)

compression_retriever = ContextualCompressionRetriever(
    base_compressor=compressor,
    base_retriever=vectorStore.as_retriever()
)

python3 query.py -q "Who started AT&T?"

Your question:
-------------
Who started AT&T?

AI Response:
-----------
AT&T - Wikipedia
"AT&T was founded as Bell Telephone Company by Alexander Graham Bell, Thomas Watson and Gardiner Greene Hubbard after Bell's patenting of the telephone in 1875."[25] "On December 30, 1899, AT&T acquired the assets of its parent American Bell Telephone, becoming the new parent company."[28]

• MongoDB Atlas
• Open AI API key
• LangChain
  • WebBaseLoader
  • RecursiveCharacterTextSplitter
  • MongoDB Atlas module
  • Contextual Compression.
  • MongoDBAtlasVectorSearch API