Vector databases are all the rage, judging by the number of startups entering the space and investors vying for a piece of the pie. The proliferation of large-scale language models (LLMs) and the generative AI (GenAI) movement have created fertile ground for vector database technologies to flourish.
Traditional relational databases like Postgres and MySQL are great for structured data (predefined data types that can be neatly filed into rows and columns), but they're great for unstructured data like images, video, email, social media, etc. Not very good for data. posts, and data that does not conform to a predefined data model.
Vector databases, on the other hand, store and process data in the form of vector embeddings, converting text, documents, images, and other data into numerical representations that capture the meaning and relationships between different data points. This is ideal for machine learning because databases store data spatially by each item's relationship to other items, making it easy to retrieve semantically similar data.
This is particularly useful for LLMs such as OpenAI's GPT-4, as AI chatbots can better understand the context of a conversation by analyzing previous similar conversations. Vector search is also useful for all kinds of real-time applications, such as content recommendations in social networks and e-commerce apps, because it allows you to see what users have searched for and instantly retrieve similar items.
Vector search can also help reduce “hallucinations” in LLM applications by providing additional information that may not have been available in the original training dataset.
“You can develop AI/ML applications without using vector similarity search, but it will require more retraining and fine-tuning,” says CEO and CEO of vector search startup Qdrant. Co-founder Andre Zayarni explained to TechCrunch. “Vector databases are useful when you have large datasets and need tools to manipulate vector embeddings in an efficient and convenient way.”
Qdrant secured $28 million in funding in January to capitalize on growth, which helped the company become one of the top 10 fastest-growing commercial open source startups last year. And it's not the only vector database startup to raise cash recently. Vespa, Weaviate, Pinecone and Chroma collectively raised $200 million last year for various Vector products.
Qdrant founding team. Image credit: Qdrant
Earlier this year, we also saw Index Ventures lead a $9.5 million seed round in Superlinked, a platform that transforms complex data into vector embeddings. And a few weeks ago, Y Combinator (YC) announced his Winter '24 cohort. This included Lantern, a startup that sells a hosted vector search engine for Postgres.
Elsewhere, Marqo raised a $4.4 million seed round late last year, quickly followed by a $12.5 million Series A round in February. The Marqo platform provides a full range of out-of-the-box vector tools for generating, storing, and retrieving vectors, allowing users to bypass third-party tools such as OpenAI and Hugging Face, all through a single API. Offers.
Marqo co-founders Tom Hamer and Jesse N. Clark previously worked in engineering at Amazon, where they focused on the “unmet need'' for semantic and flexible search across different modalities such as text and images. I realized that there was a huge need. And that's when they jumped ship to found Marqo in 2021.
“When I was working on visual search and robotics at Amazon, I was seriously considering vector search. I was thinking about new ways to discover products, and it quickly came down to vector search.” Clark told TechCrunch. “In the field of robotics, we were using multimodal search to search through many images to identify if there was something wrong with it, like a hose or a package. That would have been very difficult.”
Jesse Clark and Tom Hamer, co-founders of Marqo.Image credit: Marco
enter a company
While vector databases are a temporary stop amid all the fuss of the ChatGPT and GenAI movements, they are not a panacea for all enterprise search scenarios.
“Specialized databases tend to be completely focused on a specific use case, so they are architected with performance and user experience in mind for the tasks required, compared to general-purpose databases that must be adapted to the current design.” ” says Peter. Zaitsev, founder of database support and services company Percona, explained to TechCrunch.
Specialized databases can be superior in some respects, but this is why existing databases such as Elastic, Redis, OpenSearch, Cassandra, Oracle, and MongoDB are adding vector database search capabilities. That's why we're starting to see. Cloud service providers such as Microsoft's Azure, Amazon's AWS, and Cloudflare.
Zaitsev compares this latest trend to what happened with JSON more than a decade ago. At the time, web apps were becoming more popular and developers needed language-independent data formats that were easy for humans to read and write. In that case, new database classes appeared in the form of document databases such as MongoDB, which introduced his JSON support to existing relational databases as well.
“I think the same thing is likely to happen with vector databases,” Zaitsev told TechCrunch. “Users who are building very complex, large-scale AI applications will use a dedicated vector search database, while users who need to build a little AI functionality into an existing application will want to use a database they already use. You're more likely to use the vector search capabilities of your database.
However, Zayarni and his Qdrant colleagues believe that native solutions built around vector data offer the “speed, memory safety, and I'm confident that it will provide “scale”.
“Their argument is, 'We can also do vector searches if we want,'” Zayarni says. “Our pitch is, 'We perform advanced vector searches in the best way possible.' It's all about specialization. In reality, we recommend starting with the databases you already have in your technology stack.” If vector search is a critical component of the solution, at some point users will run into limitations.