RAG with Atlas Vector Search, LangChain, and OpenAI
Harshad Dhavale10 min read • Published Sep 18, 2024 • Updated Sep 18, 2024
With all the recent developments (and frenzy!) around generative AI, there has been a lot of focus on LLMs, in particular. However, there is also another emerging trend that many are unaware of: the rise of vector stores. Vector stores or vector databases play a crucial role in building LLM applications. This puts Atlas Vector Search in the vector store arena that has a handful of contenders.
The goal of this tutorial is to provide an overview of the key-concepts of Atlas Vector Search as a vector store, and LLMs and their limitations. We’ll also look into an upcoming paradigm that is gaining rapid adoption called "retrieval-augmented generation" (RAG). We will also briefly discuss the LangChain framework, OpenAI models, and Gradio. Finally, we will tie everything together by actually using these concepts + architecture + components in a real-world application. By the end of this tutorial, readers will leave with a high-level understanding of the aforementioned concepts, and a renewed appreciation for Atlas Vector Search!
Large language models (LLMs) are a class of deep neural network models that have been trained on vast amounts of text data, which enables them to understand and generate human-like text. LLMs have revolutionized the field of natural language processing, but they do come with certain limitations:
- Hallucinations: LLMs sometimes generate factually inaccurate or ungrounded information, a phenomenon known as “hallucinations.”
- Stale data: LLMs are trained on a static dataset that was current only up to a certain point in time. This means they might not have information about events or developments that occurred after their training data was collected.
- No access to users’ local data: LLMs don’t have access to a user’s local data or personal databases. They can only generate responses based on the knowledge they were trained on, which can limit their ability to provide personalized or context-specific responses.
- Token limits: LLMs have a maximum limit on the number of tokens (pieces of text) they can process in a single interaction. Tokens in LLMs are the basic units of text that the models process and generate. They can represent individual characters, words, subwords, or even larger linguistic units. For example, the token limit for OpenAI’s gpt-3.5-turbo is 4096.
Retrieval-augmented generation (RAG)
The retrieval-augmented generation (RAG) architecture was developed to address these issues. RAG uses vector search to retrieve relevant documents based on the input query. It then provides these retrieved documents as context to the LLM to help generate a more informed and accurate response. That is, instead of generating responses purely from patterns learned during training, RAG uses those relevant retrieved documents to help generate a more informed and accurate response. This helps address the above limitations in LLMs. Specifically:
- RAGs minimize hallucinations by grounding the model’s responses in factual information.
- By retrieving information from up-to-date sources, RAG ensures that the model’s responses reflect the most current and accurate information available.
- While RAG does not directly give LLMs access to a user’s local data, it does allow them to utilize external databases or knowledge bases, which can be updated with user-specific information.
- Also, while RAG does not increase an LLM’s token limit, it does make the model’s use of tokens more efficient by retrieving only the most relevant documents for generating a response.
This tutorial demonstrates how the RAG architecture can be leveraged with Atlas Vector Search to build a question-answering application against your own data.
The architecture of the application looks like this:
Note: For simplicity purposes, the Gradio component used for generating the application’s web-interface has not been shown in this architecture diagram.
In the following section, we will briefly discuss the different frameworks, models, and components that we will use in this tutorial, while trying to understand the nuances of some of these components.
MongoDB Atlas Vector Search allows you to perform semantic similarity searches on your data, which can be integrated with LLMs to build AI-powered applications. Data from various sources and in different formats can be represented numerically as vector embeddings. Atlas Vector Search allows you to store vector embeddings alongside your source data and metadata, leveraging the power of the document model. These vector embeddings can then be queried using an aggregation pipeline to perform fast semantic similarity search on the data, using an approximate nearest neighbors algorithm.
LangChain is a framework designed to simplify the creation of LLM applications. It provides a standard interface for chains, lots of integrations with other tools, and end-to-end chains for common applications. This allows AI developers to build LLM applications that leverage external sources of data (for example, private data sources).
Some key things to note about LangChain:
- A “chain” in LangChain is a sequence of components that can be combined together to solve a specific problem/perform a specific task.
- A chain can include components and modules such as wrappers for LLMs and vector stores, prompt templates, loaders, text splitting and chunking modules, retrievers, etc. Different components can be chained together.
- The chain takes the user's input and processes it through each component in the sequence.
This modular approach of chaining different components or modules simplifies complex application development, debugging, and maintenance. - LangChain is an open-source project launched in October 2022. The project has quickly gained popularity. It is getting a lot of adoption, including contributions from hundreds of developers on GitHub, and has an ever increasing number of integrations with external systems. It’s not only gaining a lot of adoption but also evolving rapidly to meet the needs of its growing user base.
Gradio is an open-source Python library that is used to build a web interface for ML and data science applications.
OpenAI is an AI research company and is among a handful of companies that have been instrumental in the advancement of LLMs and generative AI. OpenAI has not only developed large language models that are gaining rapid adoption, but also has developed other models that perform a variety of other tasks such as text-vectorization, image generation based on language prompt, audio-to-text conversion, etc.
When someone refers to “using OpenAI” in their application, some disambiguation is necessary because there’s no one-size-fits-all model that can be used for all use-cases. In general, different models serve different purposes. For example, OpenAI’s GPT models are language models that understand and generate languages, the DALL-E model can create images based on a textual prompt, Whisper is a model that can convert audio into text, etc.
In this tutorial, we will be using OpenAI’s embedding model and language model. Therefore, it is important to understand the distinction between the two:
Main components
Note: The setup steps as well as the Python scripts shared below can be downloaded from the Github repo.
- Install the following packages:
1 pip3 install langchain pymongo bs4 openai tiktoken gradio requests lxml argparse unstructured - Create the OpenAI API key. This requires a paid account with OpenAI, with enough credits. OpenAI API requests stop working if credit balance reaches $0.
- Save the OpenAI API key in the key_param.py file. The filename is up to you.
- Optionally, save the MongoDB URI in the file, as well.
- Create two Python scripts:
- load_data.py: This script will be used to load your documents and ingest the text and vector embeddings, in a MongoDB collection.
- extract_information.py: This script will generate the user interface and will allow you to perform question-answering against your data, using Atlas Vector Search and OpenAI.
- Import the following libraries:
1 from pymongo import MongoClient 2 from langchain.embeddings.openai import OpenAIEmbeddings 3 from langchain.vectorstores import MongoDBAtlasVectorSearch 4 from langchain.document_loaders import DirectoryLoader 5 from langchain.llms import OpenAI 6 from langchain.chains import RetrievalQA 7 import gradio as gr 8 from gradio.themes.base import Base 9 import key_param
Sample documents
In this tutorial, we will be loading three text files from a directory using the DirectoryLoader. These files should be saved to a directory named sample_files. The contents of these text files are as follows (none of these texts contain PII or CI):
- log_example.txt
1 2023-08-16T16:43:06.537+0000 I MONGOT [63528f5c2c4f78275d37902d-f5-u6-a0 BufferlessChangeStreamApplier] [63528f5c2c4f78275d37902d-f5-u6-a0 BufferlessChangeStreamApplier] Starting change stream from opTime=Timestamp{value=7267960339944178238, seconds=1692203884, inc=574}2023-08-16T16:43:06.543+0000 W MONGOT [63528f5c2c4f78275d37902d-f5-u6-a0 BufferlessChangeStreamApplier] [c.x.m.r.m.common.SchedulerQueue] cancelling queue batches for 63528f5c2c4f78275d37902d-f5-u6-a02023-08-16T16:43:06.544+0000 E MONGOT [63528f5c2c4f78275d37902d-f5-u6-a0 InitialSyncManager] [BufferlessInitialSyncManager 63528f5c2c4f78275d37902d-f5-u6-a0] Caught exception waiting for change stream events to be applied. Shutting down.com.xgen.mongot.replication.mongodb.common.InitialSyncException: com.mongodb.MongoCommandException: Command failed with error 286 (ChangeStreamHistoryLost): 'Executor error during getMore :: caused by :: Resume of change stream was not possible, as the resume point may no longer be in the oplog.' on server atlas-6keegs-shard-00-01.4bvxy.mongodb.net:27017.2023-08-16T16:43:06.545+0000 I MONGOT [indexing-lifecycle-3] [63528f5c2c4f78275d37902d-f5-u6-a0 ReplicationIndexManager] Transitioning from INITIAL_SYNC to INITIAL_SYNC_BACKOFF.2023-08-16T16:43:18.068+0000 I MONGOT [config-monitor] [c.x.m.config.provider.mms.ConfCaller] Conf call response has not changed. Last update date: 2023-08-16T16:43:18Z.2023-08-16T16:43:36.545+0000 I MONGOT [indexing-lifecycle-2] [63528f5c2c4f78275d37902d-f5-u6-a0 ReplicationIndexManager] Transitioning from INITIAL_SYNC_BACKOFF to INITIAL_SYNC. - chat_conversation.txt
1 Alfred: Hi, can you explain to me how compression works in MongoDB? Bruce: Sure! MongoDB supports compression of data at rest. It uses either zlib or snappy compression algorithms at the collection level. When data is written, MongoDB compresses and stores it compressed. When data is read, MongoDB uncompresses it before returning it. Compression reduces storage space requirements. Alfred: Interesting, that's helpful to know. Can you also tell me how indexes are stored in MongoDB? Bruce: MongoDB indexes are stored in B-trees. The internal nodes of the B-trees contain keys that point to children nodes or leaf nodes. The leaf nodes contain references to the actual documents stored in the collection. Indexes are stored in memory and also written to disk. The in-memory B-trees provide fast access for queries using the index.Alfred: Ok that makes sense. Does MongoDB compress the indexes as well?Bruce: Yes, MongoDB also compresses the index data using prefix compression. This compresses common prefixes in the index keys to save space. However, the compression is lightweight and focused on performance vs storage space. Index compression is enabled by default.Alfred: Great, that's really helpful context on how indexes are handled. One last question - when I query on a non-indexed field, how does MongoDB actually perform the scanning?Bruce: MongoDB performs a collection scan if a query does not use an index. It will scan every document in the collection in memory and on disk to select the documents that match the query. This can be resource intensive for large collections without indexes, so indexing improves query performance.Alfred: Thank you for the detailed explanations Bruce, I really appreciate you taking the time to walk through how compression and indexes work under the hood in MongoDB. Very helpful!Bruce: You're very welcome! I'm glad I could explain the technical details clearly. Feel free to reach out if you have any other MongoDB questions. - aerodynamics.txt
1 Boundary layer control, achieved using suction or blowing methods, can significantly reduce the aerodynamic drag on an aircraft's wing surface.The yaw angle of an aircraft, indicative of its side-to-side motion, is crucial for stability and is controlled primarily by the rudder.With advancements in computational fluid dynamics (CFD), engineers can accurately predict the turbulent airflow patterns around complex aircraft geometries, optimizing their design for better performance.
Loading the documents
- Set the MongoDB URI, DB, Collection Names:
1 client = MongoClient(key_param.MONGO_URI) 2 dbName = "langchain_demo" 3 collectionName = "collection_of_text_blobs" 4 collection = client[dbName][collectionName] - Initialize the DirectoryLoader:
1 loader = DirectoryLoader( './sample_files', glob="./*.txt", show_progress=True) 2 data = loader.load() - Define the OpenAI Embedding Model we want to use for the source data. The embedding model is different from the language generation model:
1 embeddings = OpenAIEmbeddings(openai_api_key=key_param.openai_api_key) - Initialize the VectorStore. Vectorise the text from the documents using the specified embedding model, and insert them into the specified MongoDB collection.
1 vectorStore = MongoDBAtlasVectorSearch.from_documents( data, embeddings, collection=collection ) - Create the following Atlas Search index on the collection, please ensure the name of your index is set to
default:
1 { 2 "fields": [{ 3 "path": "embedding", 4 "numDimensions": 1536, 5 "similarity": "cosine", 6 "type": "vector" 7 }] 8 }
Performing vector search using Atlas Vector Search
- Set the MongoDB URI, DB, and Collection Names:
1 client = MongoClient(key_param.MONGO_URI) 2 dbName = "langchain_demo" 3 collectionName = "collection_of_text_blobs" 4 collection = client[dbName][collectionName] - Define the OpenAI Embedding Model we want to use. The embedding model is different from the language generation model:
1 embeddings = OpenAIEmbeddings(openai_api_key=key_param.openai_api_key) - Initialize the Vector Store:
1 vectorStore = MongoDBAtlasVectorSearch( collection, embeddings ) - Define a function that a) performs semantic similarity search using Atlas Vector Search (note that I am including this step only to highlight the differences between output of only semantic search vs output generated with RAG architecture using RetrieverQA):
1 def query_data(query): 2 # Convert question to vector using OpenAI embeddings 3 # Perform Atlas Vector Search using Langchain's vectorStore 4 # similarity_search returns MongoDB documents most similar to the query 5 6 docs = vectorStore.similarity_search(query, K=1) 7 as_output = docs[0].page_content and, b) uses a retrieval-based augmentation to perform question-answering on the data:1 # Leveraging Atlas Vector Search paired with Langchain's QARetriever 2 3 # Define the LLM that we want to use -- note that this is the Language Generation Model and NOT an Embedding Model 4 # If it's not specified (for example like in the code below), 5 # then the default OpenAI model used in LangChain is OpenAI GPT-3.5-turbo, as of August 30, 2023 6 7 llm = OpenAI(openai_api_key=key_param.openai_api_key, temperature=0) 8 9 10 # Get VectorStoreRetriever: Specifically, Retriever for MongoDB VectorStore. 11 # Implements _get_relevant_documents which retrieves documents relevant to a query. 12 retriever = vectorStore.as_retriever() 13 14 # Load "stuff" documents chain. Stuff documents chain takes a list of documents, 15 # inserts them all into a prompt and passes that prompt to an LLM. 16 17 qa = RetrievalQA.from_chain_type(llm, chain_type="stuff", retriever=retriever) 18 19 # Execute the chain 20 21 retriever_output = qa.run(query) 22 23 24 # Return Atlas Vector Search output, and output generated using RAG Architecture 25 return as_output, retriever_output - Create a web interface for the app using Gradio:
1 with gr.Blocks(theme=Base(), title="Question Answering App using Vector Search + RAG") as demo: 2 gr.Markdown( 3 """ 4 # Question Answering App using Atlas Vector Search + RAG Architecture 5 """) 6 textbox = gr.Textbox(label="Enter your Question:") 7 with gr.Row(): 8 button = gr.Button("Submit", variant="primary") 9 with gr.Column(): 10 output1 = gr.Textbox(lines=1, max_lines=10, label="Output with just Atlas Vector Search (returns text field as is):") 11 output2 = gr.Textbox(lines=1, max_lines=10, label="Output generated by chaining Atlas Vector Search to Langchain's RetrieverQA + OpenAI LLM:") 12 13 # Call query_data function upon clicking the Submit button 14 15 button.click(query_data, textbox, outputs=[output1, output2]) 16 17 demo.launch()
The following screenshots show the outputs generated for various questions asked. Note that a purely semantic-similarity search returns the text contents of the source documents as is, while the output from the question-answering app using the RAG architecture generates precise answers to the questions asked.
Log analysis example
Chat conversation example
Sentiment analysis example
Precise answer retrieval example
In this tutorial, we have seen how to build a question-answering app to converse with your private data, using Atlas Vector Search as a vector store, while leveraging the retrieval-augmented generation architecture with LangChain and OpenAI.
Vector stores or vector databases play a crucial role in building LLM applications, and retrieval-augmented generation (RAG) is a significant advancement in the field of AI, particularly in natural language processing. By pairing these together, it is possible to build powerful AI-powered applications for various use-cases.
