What if AI agents could perform Deep Research in seconds instead of minutes?
The member of the AI:AT Coworking Hubs AI Baseline has developed a retrieval architecture for AI agents that researches and analyzes proprietary knowledge and draws conclusions from it. It delivers 44% higher response quality and operates up to 22 times faster than current agent-based retrieval systems.
Guest article by Xenia Galkina, CEO of AI Baseline
Increasingly, organizations rely on AI agents to answer questions from hundreds of thousands of pages of internal documents, reports, policies, and research. But the retrieval systems powering those agents were not built for it.
When an agent needs to find something, it searches from scratch, every time. No matter how often the same documents have been processed, the next question pays the full cost again. For instance, when an agent receives a question, it often has to run a new search, retrieve relevant passages and send them to a language model for processing.
Each step consumes paid model tokens, computing resources and time. Even when the same documents have already been processed for an earlier question, much of that work may be repeated for the next one.
Multiplied across millions of requests, that adds up to a significant amount of compute spent rediscovering what was already known. The most sophisticated versions of this approach can take up to 40+ minutes to answer a single complex question – and the fragments that answer is built from often cannot be traced to their source. As agents become more capable and more widely deployed, so does the amount of tokens, compute, and time spent repeatedly rediscovering the same information.
With current systems, two types of failures emerge:
An efficiency failure
Retrieval systems whose cost scales with every query become economically unsustainable as information production accelerates.
A transparency failure
In a world where anything can be generated, an answer without a traceable source is untrustworthy. As agents move into regulated, high-stakes domains – finance, medicine, drug discovery, compliance -reliable, traceable and verifiable outputs become an essential requirement. Agents need answers they can act on, with sources they can point to.
INFOBOX: By retrieval, we mean the process of accessing, searching, and bringing back stored information or data so an agent can use it.
Example of the question one might ask: “What decisions have we made about Project Atlas over the past year?”
Retrieval: The system searches meeting notes, emails, design documents, and reports, selects the relevant information, and provides it to the AI before it generates the answer.
What AI Baseline makes possible
AI Baseline is knowledge infrastructure for AI agents, built on a new approach to retrieval. Information is mapped into a logically connected layer built for how agents need to reason – with provenance preserved by the construction of the knowledge space.
Because the expensive work of structuring happens once, at ingestion, agentic operations become efficient and economically viable at scale. The research improves in quality, and every answer is fully traceable: you can see how the agent arrived at it.
With today’s retrieval systems, you are typically forced to choose: improving answer quality means searching across more information, making more AI calls, and waiting longer for a response. AI Baseline breaks that trade-off: per-query costs fall, workflows that were previously uneconomical become routine, and research operations run at a scale that was not previously feasible. For the first time, organizations no longer have to choose between speed, cost, and quality.
We ran the benchmarks against existing approaches: our research mode delivers 44% higher answer quality than agentic RAG, running 22× faster on 72% fewer input tokens. At maximum performance, AI Baseline Agentic Research achieves more than 2× the answer quality and remains nearly 9× faster, while costing about 15% less.
About AI Baseline
AI Baseline is a Vienna-based startup founded by Xenia Galkina, CEO, and Alexander Czech, CTO. Backed by NVIDIA Inception & AWS. Member of the AI Factory Austria AI:AT Coworking Hub since December 2025.
The founding team brings together product and engineering backgrounds with a focus on developing information retrieval systems in drug discovery, multi-agent systems for disinformation and propaganda investigations, and large-scale information infrastructure across governmental bodies and finance.
Foto: APA-Fotoservice_Krisztian Juhasz
“Working out of the AI:AT hub has been a genuine booster for us. We were one of the first startups to join the coworking space. Beyond that, through AI Factory Austria and EuroHPC we received 20,000 GPU hours on the Leonardo supercomputer, which is what made delivery of the first version of the product of this scale possible for a bootstrapping startup.”
Xenia Galkina, CEO und Co-Founder von AI Baseline
“AI Factory Austria has been a real accelerator for us. The GPU access allowed us to build a knowledge space at a scale and speed that would have been impossible for a pre-seed company on its own. Just as valuable is the community around it: being able to exchange ideas with other founders facing similar challenges, and to draw on experts across both technical and broader AI-related topics, has helped us think beyond the technology itself. For an ambitious AI startup, that combination is a genuine booster.”
Alexander Czech, CTO und Co-Founder von AI Baseline
Try it yourself
AI Baseline’s retrieval architecture system is currently available in beta for organizations that want to explore this approach with their own workflows. It is particularly relevant for technical teams deploying AI agents across their organization’s own documents and knowledge bases, including in life sciences, healthcare, legal and compliance, financial research, industrial, and public-sector environments.
Try it for free during the beta, available via API or as a Claude skill.
Working with your own documents? AI Baseline is also looking for design partners for its ingestion engine.
