Time is ticking for vector database companies.
Cloud giant AWS has launched the preview of Amazon S3 Vectors, a new vector storage solution that allows businesses to store and query AI-ready data at scale, with cost savings of up to 90% compared to traditional methods.
The service introduces a new type of bucket called vector buckets within Amazon S3, which comes with dedicated APIs to store, access, and query vector data without the need to provision any infrastructure. Each bucket can hold up to 10,000 vector indexes, and each index can store tens of millions of vectors.
“Amazon S3 Vectors is the first cloud object store with native support for large vector datasets combined with subsecond query performance,” AWS said in its blog post. “This makes it cost-effective for businesses to store AI-ready data at a massive scale.”
Vector search, commonly used in generative AI applications, involves comparing vector representations of data to find similar points based on similarity or distance metrics. These vectors are usually generated from unstructured data using embedding models.
With integrations into Amazon Bedrock, SageMaker, and OpenSearch, S3 Vectors positions itself as a versatile and cost-effective solution for AI-driven use cases like semantic search, retrieval-augmented generation (RAG), and media indexing. However, does this spell the end for specialised vector database companies like Pinecone, Weaviate, and Qdrant?
The Threat to Vector Database Companies
The announcement of S3 Vectors has sparked intense discussion, with some speculating that it could disrupt the vector database market.
Arpit Bhayani, creator of DiceDB, told AIM that while S3 Vectors will likely take some market share due to its better price-performance benefits, it won’t dominate the entire space. “They will eat the share, but not much,” he said, adding that for use cases requiring absolute performance and specialised query capabilities, users will continue preferring dedicated vector databases.
“The rise of services like Amazon S3 Vectors underscores a broader trend we’ve seen over the
past year: the decline of specialised, standalone vector databases,” Manvinder Singh, vice president of AI product management at Redis, said.
He added that pure-play vector databases are finding it increasingly difficult to evolve beyond their narrow focus.
This sentiment has also been echoed on social media. “I think S3 will end up killing more startups over the course of the next few years than AI would. S3 Tables, Express one and Vectors are going after quite a bit,” Neelesh Salian, a software engineer, wrote in a post on X.
Others were more blunt in their assessment. “S3 Vectors are here. Vector databases are dead before RAG,” Aniket Maurya, research engineer at Lightning AI, said on X.
Nicholas Khami, CEO of Trieve, stated that if he were building Trieve from scratch today, he wouldn’t choose Qdrant again. “S3 Vectors plus OpenSearch Serverless is all you need. Full stop.”
Pinecone, a leading vector database company, informed AIM that it doesn’t see the new AWS project as direct competition and reaffirmed its long-standing partnership with the cloud provider. A spokesperson noted that Pinecone has been closely collaborating with AWS for years and was named AWS’s 2024 GenAI Innovator Partner of the Year.
“We were informed of this project in advance and do not consider it to be directly competitive,” the spokesperson said, adding that Pinecone remains unmatched in agentic, search, and recommender use cases that require low latency, high QPS, and accurate retrieval at scale. “We look forward to continuing our work together to help joint customers build knowledgeable AI.”
This is not the first time vector database companies have been called out. In earlier conversations with AIM, executives from MongoDB, TiDB, and CockroachDB also argued that vector search capabilities can be integrated directly into their databases.
Boris Bialek, vice president and global field CTO at MongoDB, told AIM that vector search is simply a basic search index technology, not something magical or new. He was critical of the hype around vector databases, especially startups that branded themselves with .ai and attracted large valuations.
“Vector search is just a search technology, basically a reverse index. But companies raised massive venture capital by calling it search and adding .ai, and suddenly their valuations hit $100 billion,” he said.
Bialek pointed out that traditional databases, like Postgres with pgvector, have added vector support. However, these solutions often fail at large-scale, real-world workloads. He contrasted this with MongoDB’s approach, which integrates vector search and text search but runs them as independent, scalable processes, avoiding performance conflicts.
He added that embedding vectors into databases isn’t difficult, but doing it at scale with performance, context, and manageability is where things get challenging.
Similarly, Cockroach Labs CEO Spencer Kimball told AIM that they have integrated vector indexing into their database. However, he clarified that CockroachDB is not trying to become a general-purpose vector database or compete with OpenSearch, Elastic, or MongoDB on vector search.
“If you’re already using CockroachDB for mission-critical relational workloads, you want vector support there. Not everyone needs that. But for our users, it’s essential.”
He further said that the company is not trying to win the market for the vector index. “We’re not a vector database. However, it’s a very important modality.”
Ed Huang, CTO of PingCAP, the company behind TiDB, said that instead of building a separate system, it makes more sense to add vector indexing to an existing database that already handles other data types. He prefers storing everything in a single database with a unified interface like TiDB, rather than splitting it across specialised systems such as separate vector or document databases.
With AWS S3 Vectors, the value proposition is clear because organisations can store and query vectors directly in S3 instead of paying for a specialised vector database. This is especially relevant for cost-sensitive users, such as individual developers or small teams, who have found solutions like OpenSearch or Aurora prohibitively expensive.
Why Vector Databases aren’t Dead Yet
Specialised vector databases also offer advanced features that S3 Vectors may not fully replicate. For example, Qdrant’s recently launched Cloud Inference service simplifies AI development by generating, storing, and indexing embeddings in a single environment, reducing the complexity of managing multiple infrastructures.
Similarly, Milvus and Weaviate excel in hybrid search scenarios that combine vector and structured data queries, with sub-10ms latency on complex workloads, something S3 Vectors may struggle to match.
Bhayani said vector queries are quickly becoming standard across databases and even blob storage. “We are seeing a trend where vector queries are becoming the norm in every single database, and now even blob storage,” he said.
He added that people will initially choose convenience by using vector queries in their primary database. “Slowly, as the requirements become more niche, they will move to specialised solutions,” he concluded.
The post ‘Vector Databases are Dead Before RAG’ appeared first on Analytics India Magazine.
