Vector databases are currently all the rage in the tech world, and it isn’t just hype. Vector search has become ever more critical due to artificial intelligence advances which make use of vector embeddings. These vector embeddings are vector representations of word embeddings, sentences, or documents that provide semantic similarity for semantically close inputs by simply looking at a distance metric between the vectors.
The canonical example from word2vec in which the embedding of the word “king” was very near the resulting vector from the vectors of the words “queen”, “man”, and “woman” when arranged in the following formula:
king - man + woman ≈ queen
The fact that this works has always seemed amazing to me, but it works even for fairly large documents provided our embedding space is of sufficiently high dimension. With modern deep learning methods, you can get excellent embeddings of complex documents.
For TerminusDB we needed a way to leverage these sorts of embeddings for the following tasks that our users have asked for:
- Full-text search
- Entity resolution (finding other documents which are likely the same for deduplication)
- Similarity search (for related content or for recommender systems)
- Clustering
We decided to prototype using OpenAI’s embeddings, but in order to get the rest of the features we required a vector database.
We needed a few unusual features, including the ability to do incremental indexing, and the ability to index the basis of commits, so we know precisely what commit an index applies to. This allows us to put indexing into our CI workflows.
A versioned open-source vector database doesn’t exist in the wild. So we wrote one!
Writing a vector database
A vector database is a store of vectors with the ability to compare any two vectors using some metric. The metric can be a lot of different things such as Euclidean distance, Cosine similarity, Taxicab geometry, or really anything which obeys the triangle inequality rules required to define a metric space.
In order to make this fast, you need to have some sort of indexing structure to quickly find candidates that are already close for comparison. Otherwise, many operations will need to compare with every single thing in the database, every time.
There are many approaches to indexing vector spaces, but we went with the HNSW (Hierarchical Navigable Small World) graph. See Malkov and Yashunin. HNSW is easy to understand and provides good performance in both low and high dimensions, so is flexible. Most importantly there is a very clear open-source implementation that we found – HNSW for Rust Computer Vision.
Storing the vectors
Vectors are stored in a domain. This helps to separate different vector stores that do not need to describe the same vectors. For TerminusDB we have many different commits that all pertain to the same vectors, so it’s important that we put them all into the same domain.
.
Page
0 1 2...
———————————————————————
Vectors: | 0 [......] 2 [......]
| 1 [......] 3 [......]
The vector store is page based, where each buffer is designed to map cleanly to the operating system pages but fit the vectors we use closely. We assign each vector an index and then we can map from the index to the appropriate page and offset.
Inside the HNSW index, we refer to a LoadedVec. This ensures that the page lives in a buffer, currently loaded so we can perform our metric comparisons on the vectors of interest.
As soon as the last LoadedVec drops from a buffer, the buffer can be added back into a buffer pool to be used to load a new page.
Creating a versioned index
We build an HNSW structure for each (domain + commit) pair. If starting a new index, we start with an empty HNSW. If starting an incremental index from a previous commit, we load the old HNSW from the previous commit and then begin our indexing operations.
What is new versus what is old is all kept in TerminusDB, which knows how to find changes between commits and can submit them to the vector database indexer. The indexer only needs to know the operations it is being asked to perform (i.e. Insert, Delete, Replace).
We maintain the indexes themselves in an LRU pool that allows us to load on demand or use a cache if the index is already in memory. Since we only perform destructive operations at commits, this caching is always coherent.
When we save the index, we serialize the structure with the raw vector index as a stand-in for the LoadedVec which helps to keep the index small.
In the future, we would like to use some of the tricks we have learned in TerminusDB to keep layers of an index around, so new layers can be added without requiring each incremental index to add a duplicate when serializing. However, the indexes have proved small enough compared to the vectors we are storing that it has not mattered much.
NOTE: While we currently do incremental indexing, we have yet to implement the delete and replace operations (there are only so many hours in a week!). I’ve read the literature on HNSW and it seems that it is not yet well described.
We have a design for the delete and replace operations that we think will work well with HNSW and wanted to explain in case any technical people have ideas:
- If we are in an upper layer of the HNSW, then simply ignore the deletion – it should not matter much as most vectors are not in upper layers, and ones that are, are only for navigation
- If we are in the zero layer but not in an above layer, delete the node from the index, while trying to replace links between all neighbors of the deleted link according to closeness.
- If we are in the zero layer but also above, mark the node as deleted, and use it for navigation but do not store this node in the candidate pool.
Finding embeddings
We use OpenAI to define our embeddings, and after an indexing request is made to TerminusDB, we feed each of the documents to OpenAI which returns lists of float vectors in JSON.
It turns out that the embeddings are quite sensitive to context. We tried initially just submitting TerminusDB JSON documents and the results were not fantastic.
However, we found that if we define a GraphQL query + Handlebars template, we can create very high-quality embeddings. For People in Star Wars, this pair, which is defined in our schema, looks like this:
{
"embedding": {
"query": "query($id: ID){ People(id : $id) { birth_year, created, desc, edited, eye_color, gender, hair_colors, height, homeworld { label }, label, mass, skin_colors, species { label }, url } }",
"template": "The person's name is {{label}}.{{#if desc}} They are described with the following synopsis: {{#each desc}} *{{this}} {{/each}}.{{/if}}{{#if gender}} Their gender is {{gender}}.{{/if}}{{#if hair_colors}} They have the following hair colours: {{hair_colors}}.{{/if}}{{#if mass}} They have a mass of {{mass}}.{{/if}}{{#if skin_colors}} Their skin colours are {{skin_colors}}.{{/if}}{{#if species}} Their species is {{species.label}}.{{/if}}{{#if homeworld}} Their homeworld is {{homeworld.label}}.{{/if}}"
}
}
People object is rendered as text which helps OpenAI understand what we mean, providing much better semantics.
Ultimately it would be nice if we could guess these sentences from a combination of our schema documentation and the schema structure, which is probably also possible using AI chat! But for now, this works brilliantly and does not require much technical sophistication.
Indexing Star Wars
So what happens when we actually run this thing? Well, we tried it out on our Star Wars data product to see what would happen. First, we fire off an index request, and our indexer obtains the information from TerminusDB:
curl 'localhost:8080/index?commit=o2uq7k1mrun1vp4urktmw55962vlpto&domain=admin/star_wars'
This returns with a task-id which we can use to poll an endpoint for completion.
The index file and vector files for the domain admin/star_wars and the commit o2uq7k1mrun1vp4urktmw55962vlpto come out as:
admin%[email protected]
and
admin%2Fstar_wars.vecs
We can now ask the semantic index server about our documents at the specified commit.
curl 'localhost:8080/search?commit=o2uq7k1mrun1vp4urktmw55962vlpto&domain=admin/star_wars' -d "Who are the squid people"
We get back a number of results as JSON which look like this:
[{"id":"terminusdb:///star-wars/Species/8","distance":0.09396297}, ...]
Species/8 id:
"The species name is Mon Calamari. They have the following hair colours: none. Their skin colours are red, blue, brown, magenta. They speak the Mon Calamarian language."
Amazing! Notice that it never says squid anywhere! There is some pretty amazing work being done by our embeddings here.
Let’s have another go:
curl 'localhost:8080/search?commit=o2uq7k1mrun1vp4urktmw55962vlpto&domain=admin/star_wars' -d "Wise old man"
"The person's name is Yoda. They are described with the following synopsis: Yoda is a fictional character in the Star Wars franchise created by George Lucas, first appearing in the 1980 film The Empire Strikes Back. In the original films, he trains Luke Skywalker to fight against the Galactic Empire. In the prequel films, he serves as the Grand Master of the Jedi Order and as a high-ranking general of Clone Troopers in the Clone Wars. Following his death in Return of the Jedi at the age of 900, Yoda was the oldest living character in the Star Wars franchise in canon, until the introduction of Maz Kanata in Star Wars: The Force Awakens. Their gender is male. They have the following hair colours: white. They have a mass of 17. Their skin colours are green."
Incredible!
While we do say “oldest” in the text, we don’t say “wise” or “man”!
I hope you can see how this could be helpful for you in getting high-quality semantic indexing of your data!
Conclusion
We have also added endpoints to find neighboring documents and to find duplicates that search the entire corpus. The latter was used on some benchmarks and has performed admirably. We hope to show the results of these experiments here soon.
While there are really great vector databases out there in the wild, such as Pinecone, we want to have a sidecar that integrates well with TerminusDB and which can be used for less technical users who care about content primarily and are not going to be spinning up their own vector database.
We are really excited about the potential of this VectorLink, and would love people to have a look at what we have so far! Please forgive us a bit for the relatively sparse error handling. We’re working furiously on it!