Raiffeisenbank implemented a RAG assistant in Kotlin without Python or new expertise

Raiffeisenbank's team built a RAG assistant on Kotlin and Spring Boot — and didn't write a single line of Python. Their experience shows that the JVM stack handles AI tasks just as well as the Python stack, if the team already works in it and doesn't want to add new competencies.

Where the Task Came From

The internal systems of a large bank are layers of documents, tables, and agreements scattered across multiple places. Product owners, tech leads, CTOs, B-1 level managers, cost management employees — everyone spends time not working in the system itself, but searching for the right information: Excel files, emails, local notes, calls to colleagues. For experienced employees, this is just friction.

For new ones, it's a real barrier. The idea of an AI assistant is simple: collect answers in one place and deliver them in plain language. Not an interface for interface's sake, but real time savings of minutes on every work request.

And for new employees, the pain is especially acute — they don't yet have informal channels and don't have a person they can turn to without extra explanations.

Why JVM, Not Python

Python is the de facto standard for AI development. LangChain, LlamaIndex, HuggingFace Transformers — the entire mature AI ecosystem grew on Python and feels most at home there. It would seem like the logical choice for an AI project. But the bank's product team lives on the JVM: Kotlin, Java, Spring Boot. A transition to Python would mean bringing in new specialists, building a new stack around one product, new areas of responsibility, and additional support risks. The team chose differently — to stay in a familiar environment. Deliberately, not by inertia. Key arguments in favor of the JVM:

• Support and scalability remain within the existing team
• No need to hire Python developers or retrain current ones
• Spring Boot provides ready-made infrastructure — security, logging, deployment
• The LLM connects via API — the language model doesn't require Python on the server
• The JVM ecosystem has working libraries for embeddings and vector stores

Result: a full-fledged RAG pipeline on Kotlin, without needing to bring in new competencies for one internal tool.

Documentation as a Hidden Problem

One of the key lessons from the project: RAG works only as well as the source data. When the assistant started giving inaccurate or contradictory answers, it turned out that the reason wasn't the model or the pipeline architecture. The reason was in the documents themselves.

Scattered Excel tables with different versions of the same data, outdated instructions that no one had updated for years, duplicate entries from different sources — all this became visible immediately once the AI started relying on it. The project gave the team a reason to organize the knowledge that had previously just accumulated without structure and periodic cleanup. In essence, the AI assistant became a documentation audit tool — though no one planned it initially.

Part of the effort went not into development, but into putting the sources in order. This is an unobvious, but practically inevitable component of any RAG implementation.

What This Means

JVM teams don't have to switch to Python for AI. If expertise and infrastructure are already built — they can be used directly. A RAG pipeline on Kotlin and Spring Boot comes together without fundamental losses in functionality, but with full control within the familiar stack. For banks and fintech companies where JVM is the de facto standard, Raiffeisenbank's experience is an important precedent: AI tasks are solvable without changing tools.