Locally Deployable Enterprise Search and Q&A Solution

Organisations involved

Main Participant: Tilde is a language technology SME specialising in multilingual AI solutions. Founded in 1991, it develops neural machine translation, speech technologies, chatbots and terminology solutions for enterprises and public institutions. Tilde combines human linguistic expertise with AI-driven R&D and actively promotes digital equity for under-represented European languages through initiatives such as its open large language model, TildeOpen.

The challenge

Organisations across Europe manage large volumes of multilingual documentation, yet enterprise search and question–answering systems often fail to meet strict data sovereignty, security, and multilingual performance requirements. Cloud-based AI raises compliance concerns for regulated environments, while language models trained mainly in English perform poorly for smaller European languages such as Lithuanian.

Tilde aimed to develop a locally deployable, AI-based enterprise search and Q&A solution that guarantees full data sovereignty while delivering high-quality results in under-resourced European languages. For Tilde, the challenge was to productise advanced language technologies into a scalable platform suitable for enterprise and public-sector deployment. This required improving accuracy for Balto-Slavic languages while demonstrating secure deployment on customer infrastructure with predictable performance and cost.

Achieving competitive performance required large-scale fine-tuning and evaluation of language models beyond Tilde’s in-house capacity. FFplus support and access to HPC resources enabled distributed training and rapid optimisation, reducing systematic errors and accelerating market readiness.

The Solution

Tilde developed a modular, locally deployable enterprise assistant combining semantic search with generative question answering. The system retrieves relevant content from internal document repositories and generates context-aware answers with traceable source references. It operates entirely on on‑premises or private cloud infrastructure, ensuring compliance with GDPR and national data regulations.

High-performance computing resources provided through FFplus enabled large-scale fine-tuning of the TildeOpen language model using up to 768 GPUs. This allowed rapid experimentation, accuracy improvements for under-resourced languages and performance benchmarking, while the final solution can be deployed without ongoing HPC usage.

Impact 

The project strengthens Tilde’s market position for AI solutions in under-resourced European languages. By fine-tuning LLMs for Balto-Slavic languages, Tilde reduced error rates from over 27% to those comparable with common languages. This enables reliable deployment in regulated sectors such as government, finance, healthcare and legal services, where data sovereignty and language accuracy are critical. The solution also supports Tilde’s expansion into Central and Eastern Europe, with an expected increase of up to 150% in market size within three years.

Socially, the solution enables advanced AI services in diverse languages, and improves accessibility to trustworthy AI tools for public institutions, SMEs and end users, thereby supporting digital inclusion and access to innovation.

From an environmental perspective, local deployment reduces reliance on cloud infrastructure. HPC resources are required only during model training, while inference runs efficiently on customer-controlled infrastructure, lowering long-term energy use and carbon footprints.

Benefits

• Enables a fully on‑premises enterprise search and Q&A solution with strong data sovereignty guarantees.
• Delivers high-quality AI support for under‑resourced European languages, strengthening multilingual equity.
• Reduces decision‑making latency for end users by up to 60% through faster information retrieval.
• Provides a reusable core platform adaptable to summarisation, translation and customer support use cases.
• Lowers operational costs and environmental impact by minimising reliance on external cloud providers.