On 26 September 2025, the first in-person edition of Vector Space Day 2025, hosted by Qdrant in Berlin, brought together engineers, researchers and AI builders who are rethinking how search, retrieval and agentic workflows work today.

The event focused on three broad themes:

1. Vector search infrastructure at scale, from architecture through ingestion and indexing
2. Retrieval for agent-based workflows (agentic AI)
3. Multi-modal, hybrid and next-generation retrieval systems that go beyond classic text embeddings

Here we summarise two standout talks from our team at Superlinked, highlighting how our work ties into these themes, and provide further reading from our VectorHub content for those who want to explore deeper.

Talk 1: “Stories from the AI Search Frontier” by Daniel Svonava

‍

In this session, Daniel explored how search systems are evolving in an era of LLMs, agents and unstructured data overload. Key points included:

• The rising demand for retrieval layers that can handle unstructured data, agent workflows and real-time context rather than static pre-indexed documents.
• Why many text-only embedding plus filter systems are hitting a wall. As queries become more complex they include temporal, spatial, behavioural or structured data alongside text, and embedding these properly is critical.
• How Superlinked’s architecture supports retrieval for both human search and agent workflows, such as streaming ingestion, multi-attribute vectors and schema-aware indexing.
• Real-world use cases showing how improved retrieval can feed more capable downstream generation or agent workflows, resulting in better recall, richer candidate sets and fewer failures in multi-step reasoning.

To dive deeper into related concepts, see:

• Real Estate NLQ Agent
• Superlinked LangChain Retriever

Talk 2: “Beyond Text Only: LlamaIndex Retriever with Superlinked’s Mixture of Encoders” by Filip Makraduli

‍

Filip discussed the concept of a “mixture of encoders,” also known as encoder stacking, and how it addresses the challenge of rich, multi-attribute queries and multi-modal data. Key takeaways included:

• Modern queries often combine structured data such as numbers, categories and timestamps with unstructured text. Text embeddings alone cannot capture that complexity.
• The mixture-of-encoders approach uses separate encoders for text, numerical values, categorical metadata and temporal signals, combining them into a unified embedding space. This enables richer vectors and better recall at retrieval time, while reducing dependence on expensive reranking.
• Integration with frameworks such as LlamaIndex, and how Superlinked’s open-source library supports this method.
• Real-world outcomes: improved retrieval quality, faster iteration and better alignment between user intent and retrieved candidates, particularly in domains like e-commerce, travel and multi-faceted search.

Relevant further reading:

• Superlinked LangChain Retriever
• Superlinked on GitHub

Why It Matters

Whether you build search engines for end users, retrieval components for agent workflows, or recommendation systems that must understand both structured and unstructured inputs, the lessons from Vector Space Day 2025 matter.

• Retrieval is no longer a simple “text embedding plus top-k” problem. It must handle numeric, categorical, temporal, spatial and behavioural data.
• The candidate set before reranking matters. If you rely too heavily on reranking you may miss good candidates altogether.
• Architecture, indexing strategy, embedding model choices and schema design all come together in production retrieval systems.
• Vector search is not just “plug in an embedding model and do cosine similarity.” To scale and perform reliably, you need frameworks that manage throughput, latency, model refresh, embedding drift, versioning and metadata handling.

At Superlinked we have built our open-source framework and engineering practices to support exactly this kind of next-generation retrieval: structured and unstructured, human and agent, single-step retrieval and downstream reasoning.

‍

If the topics covered here resonate with your work, such as building retrieval for recommendations, agent workflows, multi-attribute search or real-time ingestion, we would love to help.

➡️ Book a call with us to explore your use case, constraints and how Superlinked can help you build vector-native retrieval that truly works.

Thanks to the Qdrant team and all the speakers for an inspiring event. We look forward to applying these lessons and collaborating on the next generation of retrieval systems.

‍