How to Do Semantic Search in MongoDB Using Atlas Vector Search
Rate this tutorial
Have you ever been looking for something but don’t quite have the words? Do you remember some characteristics of a movie but can’t remember the name? Have you ever been trying to get another sweatshirt just like the one you had back in the day but don’t know how to search for it? Are you using large language models, but they only know information up until 2021? Do you want it to get with the times?! Well then, vector search may be just what you’re looking for.
Vector search is a capability that allows you to do semantic search where you are searching data based on meaning. This technique employs machine learning models, often called encoders, to transform text, audio, images, or other types of data into high-dimensional vectors. These vectors capture the semantic meaning of the data, which can then be searched through to find similar content based on vectors being “near” one another in a high-dimensional space. This can be a great compliment to traditional keyword-based search techniques but is also seeing an explosion of excitement because of its relevance to augment the capabilities of large language models (LLMs) by providing ground truth outside of what the LLMs “know.” In search use cases, this allows you to find relevant results even when the exact wording isn't known. This technique can be useful in a variety of contexts, such as natural language processing and recommendation systems.
Note: As you probably already know, MongoDB Atlas has supported full-text search since 2020, allowing you to do rich text search on your MongoDB data. The core difference between vector search and text search is that vector search queries on meaning instead of explicit text and therefore can also search data beyond just text.
- Semantic understanding: Rather than searching for exact matches, vector search enables semantic searching. This means that even if the query words aren't present in the index, but the meanings of the phrases are similar, they will still be considered a match.
- Scalable: Vector search can be done on large datasets, making it perfect for use cases where you have a lot of data.
- Flexible: Different types of data, including text but also unstructured data like audio and images, can be semantically searched.
- Efficiency: By storing the vectors together with the original data, you avoid the need to sync data between your application database and your vector store at both query and write time.
- Consistency: Storing the vectors with the data ensures that the vectors are always associated with the correct data. This can be important in situations where the vector generation process might change over time. By storing the vectors, you can be sure that you always have the correct vector for a given piece of data.
- Simplicity: Storing vectors with the data simplifies the overall architecture of your application. You don't need to maintain a separate service or database for the vectors, reducing the complexity and potential points of failure in your system.
- Scalability: With the power of MongoDB Atlas, vector search on MongoDB scales horizontally and vertically, allowing you to power the most demanding workloads.
Now, let's get into setting up a MongoDB Atlas cluster, which we will use to store our embeddings.
Step 1: Create an account
Step 2: Build a new cluster
After creating an account, you'll be directed to the MongoDB Atlas dashboard. You can create a cluster in the dashboard, or using our public API, CLI, or Terraform provider. To do this in the dashboard, click on “Create Cluster,” and then choose the shared clusters option. We suggest creating an M0 tier cluster.
Step 3: Create your collections
Now, we’re going to create your collections in the cluster so that we can insert our data. They need to be created now so that you can create an Atlas trigger that will target them.
For this tutorial, you can create your own collection if you have data to use. If you’d like to use our , you need to first create an empty collection in the cluster so that we can set up the trigger to embed them as they are inserted. Go ahead and create a “sample_mflix” database and “movies” collection now using the UI, if you’d like to use our sample data.
To proceed to the next step using OpenAI, you need to have set up an account on OpenAI and created an API key.
If you don't want to embed all the data in the collection you can use the "sample_mflix.embedded_movies" collection for this which already has embeddings generated by Open AI, and just create an index and run Vector Search queries.
Step 1: Create a trigger
To create a trigger, navigate to the “Triggers” section in the MongoDB Atlas dashboard, and click on “Add Trigger.”
Step 2: Set up secrets and values for your OpenAI credentials
Go over to “App Services” and select your “Triggers” application.
Create a new Value
Select “Secret” and then paste in your OpenAI API key.
Then, create another value — this time, a “Value” — and link it to your secret. This is how you will securely reference this API key in your trigger.
Now, you can go back to the “Data Services” tab and into the triggers menu. If the trigger you created earlier does not show up, just add a new trigger. It will be able to utilize the values you set up in App Services earlier.
Step 3: Configure the trigger
Select the “Database” type for your trigger. Then, link the source cluster and set the “Trigger Source Details” to be the Database and Collection to watch for changes. For this tutorial, we are using the “sample_mflix” database and the “movies” collection. Set the Operation Type to 'Insert' ‘Update’ ‘Replace’ operation. Check the “Full Document” flag and in the Event Type, choose “Function.”
In the Function Editor, use the code snippet below, replacing DB Name and Collection Name with the database and collection names you’d like to use, respectively.
This trigger will see when a new document is created or updated in this collection. Once that happens, it will make a call to the OpenAI API to create an embedding of the desired field, and then it will insert that vector embedding into the document with a new field name.
Now, head over to Atlas Search and create an index. Use the JSON index definition and insert the following, replacing the embedding field name with the field of your choice. If you are using the sample_mflix database, it should be “plot_embedding”, and give it a name. I’ve used “moviesPlotIndex” for my setup with the sample data.
First, click the “search” tab on your cluster
Then, click “Create Search Index.”
Create “JSON Editor.”
Then, select your Database and Collection on the left and a drop in the code snippet below for your index definition.
Now, you need to insert your data. As your data is inserted, it will be embedded using the script and then indexed using the KNN index we just set.
If you’re going to use the sample movie data, you can just go to the cluster, click the … menu, and load the sample data. If everything has been set up correctly, the sample_mflix database and movies collections will have the plot embeddings created on the “plot” field and added to a new “plot_embeddings” field.
Once the documents in your collection have their embeddings generated, you can perform a query. But because this is using vector search, your query needs to be transformed into an embedding. This is an example script of how you could add a function to get both an embedding of the query and a function to use that embedding inside of your application.
This script first transforms your query into an embedding using the OpenAI API, and then queries your MongoDB cluster for documents with similar embeddings.
Support for the '$vectorSearch' aggregation pipeline stage is available with MongoDB Atlas 6.0.11 and 7.0.2.
Remember to replace 'your_openai_key', 'your_mongodb_url', 'your_query', ‘<DB_NAME>’, and ‘<COLLECTION_NAME>’ with your actual OpenAI key, MongoDB URL, query, database name, and collection name, respectively.
And that's it! You've successfully set up a MongoDB Atlas cluster and Atlas trigger which calls the OpenAI API to embed documents when they get inserted into the cluster, and you’ve performed a vector search query.
If you prefer learning by watching, check out the video version of this article!