The Fast-Emerging World of Vector Databases

AWS, Google Cloud, Microsoft, MongoDB, and others are racing to deliver vector capabilities for AI development and applications.

Aug 05, 2023

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist who went on to work in strategic communications at Oracle, IBM, MongoDB, and now Method Communications. If you received this newsletter, you’re a subscriber (thank you!) or someone forwarded it to you. Subscriptions are free.

Another day, another vector database. The tech industry is being taken over by vector databases—an arcane technology even in the complex world of database software. But vector databases and vector search are very quickly becoming mainstream, driven by—you guessed it—the rush to build AI applications, including but not limited to generative AI.

Vector databases store, search, and retrieve vectors, which are long strings of numbers representing documents, images, and other data types. Use cases include recommendations, personalization, similarity search (text, image, audio), and “long-term memory” for Large Language Models.

Vector technologies have been around for years as a special-purpose technology. Facebook engineers released Facebook AI Similarity Search (FAISS) back in 2017.

Now vector capabilities are popping up everywhere as database vendors rush to become part of the AI tech stacks that IT and dev teams use to move from AI pilot projects to enterprise deployment.

In some cases, these are full-fledged vector database management systems. In others, they’re vector capabilities added to widely used DBMSes such as PostgreSQL.

I’ve been writing about vector databases in the Cloud Database Report for more than two years. In fact, my first podcast was with Edo Liberty, CEO of Pinecone Systems, an early leader in vectors. There’s a link to that podcast at the bottom of this blog post.

Now, everyone is jumping on the bandwagon. AWS, Google Cloud, Microsoft, and MongoDB are all getting in on the action.

Vector startups include Chroma, LanceDB, Pinecone Systems, Qdrant, Weaviate, and Zilliz.

Following is my recap, in chronological order, of 12 noteworthy developments over the past 15 weeks.

Timeline of vector announcements (2023)

April 21: Weaviate raised $50 million in Series B funding, led by Index Ventures. The startup describes its platform as an “AI-native” vector database.
April 26: Pinecone secured $100 million in Series B funding, led by Andreessen Horowitz. At the time, Pinecone was valued at $750 million, just two years after its vector database was first introduced. CEO Edo Liberty said “the rise of generative AI propelled Pinecone to become an integral part of the software stack for AI applications.”
June 14: Zilliz, a startup with $113 million in funding , announced a free tier of its vector database, available as a cloud service. Zilliz is the lead developer of Milvus, an open source vector database. [An interesting bit of trivia: Zilliz is a neologism (i.e. made-up word) and a palindrome that stands for zillions of zillions. “We use it to show how much unstructured data we are trying to sort through,” founder and CEO Charles Xie told me. See “Database Startups X, Y, Z.”]
June 18: Microsoft launched, in preview mode, vector search for Azure Cognitive Search, its AI-powered search and retrieval service. Microsoft describes it as a contemporary retrieval system for LLMs and generative AI apps. Azure Cognitive Search ties in with Microsoft’s Azure AI and Azure OpenAI services, which means customers can use it with ChatGPT.
June 22: MongoDB announced Atlas Vector Search, a vector search capability for the company’s Atlas database services platform. Use cases include similarity search, recommendation, Q&A systems, personalization, and long-term memory for LLMs. Here’s a short explainer video from MongoDB.
June 26: Google Cloud added vector support to two of its PostgreSQL database services—Cloud SQL and AlloyDB. The new capability employs pgvector, an open-source vector extension for Postgres. Says Google Cloud, “Vector support in PostgreSQL means you can use your existing operational database to power AI-enabled experiences and leverage your existing PostgreSQL skills including everything that the PostgreSQL ecosystem has to offer.”
July 6: Tencent unveiled a vector database called Tencent Cloud VectorDB, which according to one report is already being used in several of the company’s services.
July 18: DataStax introduced vector search for its Astra DB database service, which is based on the open source Apache Cassandra NoSQL database. Generative AI use cases for vector search in Astra DB/Cassandra include answering questions with natural language processing (NLP), semantic or similarity search, and semantic caching, which is important for performance as the number of API calls to LLMs scales.
July 25: Alibaba adds vector engine to its AnalyticDB database, according to CloudTech News. There are more details on the Alibaba website here.
July 26: AWS announced a preview release of vector capabilities for Amazon OpenSearch Serverless, its version of the popular Elasticsearch search engine. AWS says the new vector capabilities can be used to create ML-enabled search and generative AI apps that perform natural language searches and queries.
Aug. 1: Database startup Neon raised $46 million in Series B funding, bringing its total to $104 million, with Databricks and Snowflake among the investors. Neon developed its own vector extension, pg_embedding. “It uses one of the more modern algorithms, so it’s a lot faster” than pgvector, CEO Nikita Shamgunov told VentureBeat.
Aug 1: TileDB added vector search to its array database. As a result, says CEO Stavros Papadopoulos, TileDB manages the vector embeddings along with the raw original data (e.g., images, text files), the ML embedding models, and the other data modalities in app (tables, genomics, point clouds).

Further info

As you can see, there’s a tremendous amount of activity around vector capabilities, and no doubt there will be more. And the above list is not exhaustive; other database vendors support vector search and embeddings, too. I will continue to report on this fast-changing area of the market.

Finally, here are a few additional resources:

A detailed, explanatory blog post by TileDB CEO Stavros Papadopoulos
My June 2021 podcast with Pinecone CEO Edo Liberty

John Foley is VP of Content and Thought Leadership with Method Communications. The Cloud Database Report is independently published and unaffiliated with Method, and the views here are my own. Connect with me on LinkedIn.

Share Cloud Database Report