Vector databases as the foundation for AI applications

An AI assistant that answers questions about internal company documentation, an agent that reviews software code against a proprietary knowledge base, a search engine that understands meaning rather than matching keywords: all these applications share one technical requirement. They need a way to store large volumes of information so it can be searched quickly and semantically. Vector databases fulfil exactly that role.

Amsterdam-based Weaviate, founded in 2019 and now backed by more than €45 million in a Series B round, builds open-source vector database infrastructure for billion-scale applications. The company represents an infrastructure layer that is becoming increasingly visible as more startups build AI products and need memory mechanisms that go beyond a conventional relational database.

Choosing the right memory infrastructure has a direct bearing on how well an AI application performs in production. That makes it a strategic decision for any founder building an application on large language models.

What a vector database does and why it matters

Traditional databases store structured data and search by exact matches. A vector database works differently: text, images or other data are converted into mathematical representations, known as embeddings, that capture the meaning of the content. Two sentences that mean the same thing but use different words will end up close together in the vector space.

That property makes vector databases essential for retrieval-augmented generation, in which a language model does not rely solely on its training data but also retrieves relevant passages from an external knowledge source. A RAG system first fetches relevant pieces of text from the vector database and passes them as context to the language model. The model can then answer questions about documentation that was updated after its training period, or about internal company information it has never seen.

Weaviate supports this pattern and adds modular integrations for popular embedding models and search methods. The open-source approach makes it possible to self-host the database, which offers a practical advantage for applications involving sensitive business data, in addition to licensing cost benefits.

Memory as a building block for AI agents

Beyond RAG applications, the importance of vector databases is growing for AI agents: systems that autonomously take steps to complete a task and must retain context across sessions. An agent handling multiple tasks needs to know what was discussed earlier, which decisions have already been made, and which information is relevant for the next step.

This requires several types of memory. Short-term memory typically resides in the language model's own context window, but that context is finite. External storage systems are needed for long-term memory, stored experiences and searchable knowledge bases. In that architecture, vector databases take on the role of semantic memory.

Dutch company General Intuition, founded in 2025 and backed by €114 million in a seed round, trains AI agents on billions of gameplay frames for spatial reasoning. That type of training produces large volumes of experience data that must be efficiently searchable. It illustrates how memory infrastructure plays a role not only in text-based applications but also in agents operating in simulated environments.

What application startups weigh when choosing memory infrastructure

For a startup building an AI product, the choice of memory infrastructure is both practical and financial. Fully managed cloud solutions reduce operational burden but carry per-query costs that can escalate quickly at scale. Self-hosted open-source solutions such as Weaviate give more control over data and costs, but require internal expertise in infrastructure management.

Compliance is another factor. Companies working with personal data or commercially sensitive information must be able to demonstrate where data is stored and how access is governed. Deeploy, the Rotterdam-based AI governance platform that closed an €8.5 million Series A round in 2024, addresses precisely this question: how to keep AI systems auditable and compliant in a production environment. The memory layer is one of the components that demands attention in that context.

A third consideration is the quality of the training data that feeds the embeddings. Toloka, headquartered at Schiphol and backed by €64 million in a Series B round earlier this year, delivers high-quality training data through a combination of human experts and automation. How well a vector database performs depends in part on the quality of the embeddings, which in turn depends on the model used to generate them and the data on which that model was trained.

Infrastructure versus the application layer

The rise of vector databases sharpens the distinction between infrastructure startups and application startups. Weaviate provides a building block; the value lies in what others build with it. CuspAI in Amsterdam, which has raised €111 million in a Series A and uses AI to search for and generate new materials, is an example of a deep-tech application that relies on this infrastructure layer to search materials data rapidly.

For investors, this presents a choice between two risk profiles. Infrastructure-layer companies such as Weaviate have potentially broad adoption, but their margin model competes with large cloud providers that incorporate comparable services into their own platforms. Application startups bear the risk of product-market fit in a specific segment, but can achieve higher margins if that fit holds.

Within the Dutch and broader European AI ecosystem, the presence of infrastructure companies such as Weaviate is relevant to the ecosystem's overall productivity. Founders building applications can opt for a European infrastructure player with open-source roots, which offers advantages in terms of both data governance and vendor independence. As EU regulation around AI and data continues to tighten, that consideration may carry increasing weight in the product decisions of startups based here.