FAQ

How does K Pro protect confidential data?

For each K Pro user, only the chat history is saved in the database and it is always associated with the given user + their organization. Only an authenticated user can access their own chat history.

The database itself is a managed RDS instance on Owkin’s AWS account and only accessible to the backend service that needs it via a secure service account + network policies.

Can I access the codebase and decision making process of K Pro?

Owkin understands the value and importance of interpretability with ML and AI tools. While the codebase of K Pro is proprietary and confidential to Owkin, at runtime, users can inspect the reasoning process undertaken by agents in the system. Where plots are generated, or a coding agent generates code for a complex task, this will be available to the user.

How does Owkin standardize and harmonize diverse datasets and turn them into a productized agentic system?

On top of the diverse datasets, K Pro agentic system is built over tools and agents which take into account the diversity of datasets. The K Pro orchestration allows to capture intent from user questions and to call the right tools in order to find proofpoints across natural language scientific reports (e.g. in literature or patents), structured gene databases (e.g. OpenTargets) or multi-omics patient cohorts. Standardization / harmonization is performed by dedicated transformations optimized for the usage by the tools and agents in K Pro agentic system.

How does Owkin ensure a proper and deterministic retrieval of citations and references?

We use RAG ( Retrieval-Augmented Generation) on pubmed abstracts and we make sure the pubmed ids returned do exist. There is still the possibility of the LLM giving an answer that doesnt match the content of the retrieved pubmed IDs but it is low. We also evaluate internally against public benchmarks related to literature review. We also know we can improve our daily evaluation tests to analyze the actual answer uses the retrieved pubmed IDs.

How is K Pro different from other LLMs?

Owkin K is purpose-built for biomedical research, offering unique features and scientific rigor that go far beyond the capabilities of general-purpose LLMs. Here’s what sets K apart:

1. Smarter Scientific Reasoning

Unlike generic LLMs that often provide surface-level or generic answers, K delivers scientifically robust reasoning. It supports every analysis with statistical evidence—such as p-values and population-level data—so you can trust the validity of its findings and never have to rely on guesswork.

2. Exclusive Access to Cutting-Edge Data

Owkin K provides access to specialized datasets unavailable in mainstream LLMs. For example, you can analyze MOSAIC Window, a subset of the largest multiomic spatial oncology dataset, unlocking insights from a unique and powerful data resource.

3. Interactive, User-Driven Visualizations

With K, data exploration is intuitive and interactive. You can iterate on data visualizations in real time using plain language—just ask to filter charts by new variables, compare groups, or adjust parameters. This makes deep data analysis both faster and more engaging.

4. Publication-Ready Scientific Writing

K generates responses that sound like real scientific writing. Its outputs follow accepted scientific conventions, producing clear, coherent, and publication-ready paragraphs—no more generic or off-topic content.

5. Designed by Scientists, for Scientists

While generic LLMs are built for broad, everyday tasks (like writing poems or planning trips), K is laser-focused on the needs of researchers. It’s engineered by scientists to fit the workflows, language, and standards of the biomedical community, ensuring relevance and credibility in every interaction.

Owkin K empowers researchers with rigorous scientific reasoning, exclusive data access, interactive analytics, and precise scientific communication—delivering specialized support that generic LLMs simply can’t match.

How does Owkin prevent bias in K Pro recommendations?

We conduct bias audits during model development and testing. Our models are evaluated across diverse demographic datasets, and we partner with academic and clinical collaborators to validate performance in real world settings.

At Owkin, preventing bias in K Pro's recommendations is a central focus throughout the model development lifecycle. We conduct comprehensive bias audits during both development and testing phases, using a variety of statistical and qualitative checks to identify and mitigate potential sources of bias.

Our models are systematically evaluated across diverse demographic and clinical datasets to ensure that recommendations are robust and equitable for all patient groups. To further strengthen our safeguards, we collaborate closely with leading academic and clinical partners who independently assess K Pro's performance in real-world settings. Their feedback helps us refine our models and address any disparities that might emerge.

Additionally, we maintain ongoing monitoring after deployment, so that any new or unforeseen sources of bias can be quickly identified and corrected. Our commitment to transparency means that users can always trace the data sources and rationale behind K Pro’s recommendations.

By combining rigorous technical evaluation with diverse, real-world validation and continuous monitoring, we strive to deliver recommendations that are fair, reliable, and trustworthy for everyone.

How does K Pro guard against IP leakage?

All data in K Pro is segregated by customer to ensure confidentiality. Access to data by Owkin employees is limited to those who have an operational role requiring maintenance access. System integrity and information security is maintained through multiple layers including 24/7 monitoring.

Can anyone see the data uploaded to K Pro?

No. Data uploaded to K Pro can only be seen by you (the original user) and other members of your organization. Your data is never shared with or visible to users outside your organization.

How does Owkin measure and prevent hallucinations?

We have daily monitoring in place (e.g., using a metric like Tool Call Accuracy (TCA)) to track the percentage of times the correct tool is identified and called. This helps us quickly identify systemic issues where an agent might be "hallucinating" the need for a specific tool when another is more suitable, or failing to identify any tool when one is needed. Parameter Monitoring: Once a tool is called, it's important that the correct parameters are passed to it. Incorrect parameters can lead to wrong actions, all of which manifest as hallucinations in the agent's final response. We monitor the accuracy and completeness of parameters passed to tools via some automated tests on a set of evaluation questions.

How would different LLMs or versions of the same LLM affect final outcomes?

The final outcomes would be affected by the ability of the LLMs to understand the user questions and call the right tools to answer it. Usually, more recent versions of LLMs have better performance on a tool call task as shown here.

Also, switching from one LLM to another means doing the necessary changes in prompting and context engineering to reach the best results possible for a given LLM (to be assessed via our evaluation automations).

Who owns the output from user prompts in K Pro?

Owkin does not claim rights over user submitted content or outputs. Owkin will only benefit from a license to use the content and output to improve its products and services, in particular to improve users exeperiences. Owkin is committed to protect academic freedom.

How to make sure that the scientific conclusions generated by K Pro are evidence-based and accountable?

At Owkin, we recognize that any scientific tool—whether it’s AI-driven, a spreadsheet, or statistical software—can be misused or lead to misleading conclusions if not applied responsibly. That’s why K Pro is built with robust safeguards to promote transparency, accountability, and scientific rigor.

Every result generated by K Pro is grounded in evidence from authoritative sources such as PubMed and validated biological knowledge bases. For each output, users can trace back recommendations to their original sources, ensuring full provenance and transparency. Our explainability features log every data source, model decision, and rationale, making the reasoning process fully visible and auditable.

To further reduce the risk of false or hallucinated information, we’ve implemented technical checks to verify that every cited source exists and is relevant to the scientific question at hand. However, we acknowledge that no AI system is infallible—scientific oversight and expert review remain essential. K Pro is designed to assist researchers, not replace human expertise or decision-making.

With these measures, we empower users to make informed, reliable decisions while maintaining the highest standards of scientific integrity.

Does Owkin own its full end to end tech stack? If not, how does Owkin vet their vendors?

At Owkin, keeping your data secure is our highest priority. While much of our technology is developed and managed in-house, we also partner with select, highly reputable vendors who must meet our stringent privacy, security, and ethics standards. Each partner is carefully vetted through rigorous due diligence, including detailed security assessments and contractual requirements aligned with our own commitments.

To ensure the highest standards of information protection, we employ robust organizational and technical measures, conduct regular internal and external audits, and perform comprehensive Security Risk Assessments with every major change to our systems. When integrating large language models or other third-party components, we choose hosting options that guarantee privacy and confidentiality for all data and outputs. This privacy-first approach ensures full compliance with GDPR and HIPAA requirements.

Owkin is certified to ISO 27001:2022 for information security and ISO 13485:2016 for medical device quality, reflecting our ongoing dedication to safeguarding your data. With these measures in place, you can be confident that your information is protected at every stage.