Unlocking knowledge with AI
AI hub “sparkie”
Services
AI Engineering & Compound AI Systems
Web Development
Cloud Development
Technologies
RAG (Query Expansion, HyQE, Vektor-Datenbanken)
Langchain
LangGraph
Transcription (STT)
Infrastructure as Code (IaC)
Python
React
From AI hype to AI strategy
Using generative AI in a secure and compliant way
When the hype around generative AI took off with the release of ChatGPT 3.5 in late 2022, slashwhy responded immediately: An interdisciplinary task force analyzed the opportunities and risks of the emerging technology, aiming to identify meaningful use cases – both internally and for future client projects.
As a software service provider, slashwhy already had solid experience in the field of artificial intelligence, particularly with classical machine learning methods. But open access to powerful language models opened up a whole new dimension: Generative AI suddenly became broadly accessible and highly versatile.
However, it quickly became clear that, particularly in the context of projects, legal challenges – such as data protection, uncertainties regarding rights of use for AI-generated output, or liability issues – pose major hurdles to the use of external tools. To enable experimentation despite these challenges, a self-hosted solution was needed.
That’s how sparkie was born: a strategically driven platform idea, developed from early exploration efforts.
sparkie as an internal AI hub
Scalable AI platform for exploration and productive use
Driven by the goal of using generative AI at slashwhy in a secure, transparent, and practical way, a clear product vision quickly emerged: sparkie was to become an internal AI hub that consolidates key AI functions while meeting technical, legal, and organizational requirements.
Another important impulse came from everyday work: The inadequate search function of the internal knowledge platform– based on a widely used standard solution – made it difficult for employees to access relevant information and clearly highlighted the potential of AI-powered solutions.
At the same time, sparkie was designed to provide space for experimentation and hands-on learning – offering a safe environment where developers could deepen their skills in prompt engineering, vector databases, and semantic search. In parallel, a company-wide AI policy was developed to define clear guidelines for the responsible use of generative AI tools.
As a software company, it was clear to us: if we want to use generative AI internally, it has to be secure, compliant, and transparent. That’s why we deliberately chose a self-hosted solution in our private cloud.”
Dr. Joachim Wilharm
Managing Director at slashwhy
As developers, we didn’t just want to use generative AI – we wanted to understand it. That’s why we built everything ourselves: from the RAG architecture and knowledge preparation to the full retrieval pipeline and semantic search.”
Tobias Busch
Software Engineer at slashwhy
We firmly believe that the dialog-based use of internal knowledge holds tremendous potential for many companies – from chatting with machine data to knowledge assistants for support teams. sparkie is a strong example of how generative AI can unlock internal knowledge in a new way.”
Dr. Joachim Wilharm
Managing Director at slashwhy
sparkie gave us the opportunity to build extensive hands-on expertise – from integrating large language models to optimizing RAG pipelines and addressing legal challenges. A big advantage was being able to work entirely on local hardware at first.”
Philipp Hebing
Software Engineer at slashwhy
With our internal AI hub sparkie, we demonstrate how generative AI can be used responsibly and purposefully in a business context – with a clear use case and a deep understanding of how AI systems work. That’s exactly what our clients expect from us.”
Dr. Joachim Wilharm
Managing Director bei slashwhy
AI engineering in-house
Deep expertise in LLMs and AI system
architecture
From the very beginning, building in-depth expertise in AI engineering was a central goal for slashwhy. The company therefore deliberately opted for an in-house development. The objective was not only to apply new technologies but to truly understand them at their core and evolve them internally.
It all started with a prototype for a transcription tool, developed during an internal hackathon. A dedicated core team of two developers then took over the implementation – supported by additional colleagues who contributed temporarily alongside their client projects.
To support continuous improvement, alpha and beta user groups were intentionally involved. Simple feedback mechanisms helped the team capture usage patterns, prioritize input, and refine features accordingly.
As the technical foundation, an LLM (Large Language Model) was initially hosted on existing on-premise hardware and later migrated to a private Azure cloud infrastructure to ensure performance, scalability, and stable production operations.
Use Case
AI-based knowledge chat
From static search to conversational knowledge platform
Many companies rely on standard intranet solutions for internal knowledge management, but their search functions are often inefficient and unhelpful. Instead of relevant results, users are frequently presented with long lists – where even pages with matching keywords or titles appear far down the ranking. The reason is usually a lack of semantic understanding: these systems don't recognize content relationships and rank results solely based on keyword patterns.
At slashwhy, this was the starting point: the need to access information more quickly, in context, and across departments led to the development of the knowledge chat feature in sparkie.
Its technical foundation is a Retrieval-Augmented Generation (RAG) architecture, which connects LLMs to a vector database and enables conversational, semantic search.
Retrieval-Augmented Generation,
Query Expansion & HyQE
The solution at a glance:
Rule-based data processing
Content with varying formats – from plain text and tables to bullet points – is automatically converted into a structure that can be processed by LLMs.Intelligent data import
When transferring content from the standard intranet solution into the vector database, the system uses HyQE (Hypothetical Query Embeddings) to generate hypothetical questions for each paragraph – creating a semantically enriched knowledge structure.UX-focused query handling
sparkie automatically expands each user prompt into several alternative prompts. With query expansion, the system improves search result quality even when users use vague language or miss key technical terms.
The outcome:
Fast access to relevant knowledge
Cross-functional summaries
Context-aware answers instead of static search results
This use case demonstrates the potential of generative AI far beyond intranet search. Machine data, technical documentation, product knowledge, or support content can also be transformed into a powerful, conversational knowledge platform using a similar approach.
Building know-how &
learnings
Lessons learned in AI engineering
The development of sparkie was not just a technical endeavor for slashwhy – it was an intensive learning process. The AI team built deep expertise in working with generative AI, vector databases, and retrieval systems – both in terms of technical implementation and conceptual design. One key takeaway: the quality of AI-generated results is directly dependent on the quality of the underlying knowledge base. To make unstructured content usable with generative AI, targeted preparation is essential. Information must be structured in a way that is processable and semantically accessible to language models.
At the same time, the project helped strengthen AI literacy across the entire organization, including a responsible approach aligned with the principle of Human in Command. During the rollout of sparkie, all employees were onboarded and sensitized to the responsible use of generative AI. Internal formats such as the "AI Café" now provide ongoing space for knowledge sharing, questions, and cross-team exchange – even beyond the development teams.
sparkie grows with new data sources
Next stage of development: knowledge chat with structured data
With sparkie, slashwhy has created an internal AI platform designed for targeted growth and scalability. In addition to UX improvements, the next step will be integrating further internal knowledge sources. The connection of the database-driven people/project matching platform /skills – also developed in-house – is already underway. What’s already clear: connecting structured data sources to an LLM poses fundamentally different technical and methodological challenges than processing unstructured content.
Looking ahead, relevant knowledge from various systems is set to be centrally consolidated in an internal data warehouse and made accessible via sparkie – context-aware, conversational, and secure. The goal remains to continue expanding sparkie as the central AI hub, making business-critical knowledge securely, contextually, and dialog-based accessible – and to enable additional use cases in the process.
Achievement & Outcome
Established a central AI hub: sparkie brings together key AI capabilities in a scalable platform for the secure and productive use of generative AI within the company.
Data privacy-compliant access to AI: The platform enables the use of powerful language models without exposing sensitive data to external services.
Made internal knowledge conversationally accessible: With sparkie, company-wide knowledge is now securely accessible via chat – context-aware and user-friendly.
Implemented scalable cloud infrastructure: Hosting in a private Azure cloud ensures high performance, flexibility, and stable operations.
Effectively prevented shadow IT: sparkie provides a secure internal alternative to unregulated external AI tools.
Built AI engineering expertise: The team developed deep know-how in working with LLMs, vector databases, and semantic search.
Laid the foundation for client projects: The internal platform now serves as a technological and methodological base for future AI applications in customer contexts.
FAQ
A Large Language Model (LLM) is an AI model designed to understand and process natural language. LLMs are trained on massive amounts of text data, enabling them to generate text, answer questions, create summaries, and perform translations. Well-known examples include GPT-4 and Llama. LLMs are widely used in applications such as chatbots, search engines, and automation solutions.
Retrieval-Augmented Generation (RAG) is an AI architecture that combines a language model (LLM) with an external knowledge source, such as a vector database. In the first step (Retrieval), relevant text segments (chunks) are identified from a database or document collection based on the user’s query. In the second step (Augmentation), the user query is enriched with the retrieved information. In the final step (Generation), the LLM combines the original question with the provided context to generate a precise and well-founded answer.
By incorporating up-to-date and domain-specific knowledge, RAG delivers significantly more relevant results than standalone language models without access to external data.
Large Language Models (LLMs) cannot process an entire knowledge base at once. Instead, they work best with smaller text segments known as chunks. This segmentation is important for three main reasons:
Efficiency: Instead of scanning entire documents, the system retrieves only the most relevant text segments for a query.
Precision (RAG): In Retrieval-Augmented Generation, the model receives exactly these relevant chunks as context. This reduces computational effort and leads to faster, more accurate responses.
Context preservation: Chunks must be structured meaningfully. Only when contextual relationships are maintained within a chunk can the AI generate high-quality and accurate responses.
Chunking enables AI systems to access extensive knowledge bases without being overwhelmed by large volumes of data.
Query Expansion is a method used in generative AI systems, particularly in dialog-based access to specialized knowledge sources. The original user query is automatically expanded into additional alternative prompts. To improve contextual understanding, semantic information such as synonyms, paraphrases, and related technical terms is systematically incorporated during prompt generation.
The goal is to retrieve more relevant content – even if users do not use the exact terminology. Query Expansion therefore increases retrieval accuracy, especially for complex or ambiguous queries.
HyQE (Hypothetical Query Embeddings) is an approach used in semantic search systems powered by AI. The AI first analyzes all segmented documents (chunks) and generates potential questions that each section could answer. These hypothetical questions are stored as vector embeddings.
During future searches, the system can retrieve more relevant and comprehensive results because it has already anticipated potential user queries. This significantly improves discoverability and aligns search results more closely with user intent.
“Human in Command” refers to a principle for the responsible use of generative AI. While AI systems can assist with tasks – such as automation or content generation – the final decision, validation, and responsibility remain with a human. This principle protects against so-called “AI slop” – the uncritical acceptance of AI-generated content that may be incorrect, incomplete, or biased. Human in Command ensures that AI applications are used in a controlled, transparent, and responsible manner.
An internal AI chat provides organizations with a secure and data privacy-compliant alternative to public tools like ChatGPT or Gemini. While such internal solutions still rely on existing Large Language Models (LLMs), they are operated within the company’s own infrastructure – either on-premise or in a private cloud. This ensures that sensitive data remains fully under the organization’s control.
At the same time, internal AI systems can be tailored to company-specific requirements, such as integrating proprietary knowledge sources, implementing custom access controls, or optimizing the user experience.
Looking for answers?
Are you looking for a partner for a collaborative project or would you like to hear our expert opinion? Feel free to send us an email or find more contact information in our contact section.
