AI to the point

The technology behind AI

Artificial intelligence consists of a large number of technological components and principles. In the following, we provide a compact overview of the most important concepts, architectures and methods that make modern AI systems possible. From language models and explainability to data protection aspects.

The following overview provides an introduction; developments in the field of AI are dynamic and constantly changing. Would you like to find out how these technologies can be used in your organization? Get in touch with us.

Basics

Modern AI is based on various technologies. We present the most important concepts in the following sections.

Generative AI

Generative AI describes a new form of artificial intelligence (AI) that not only analyzes content, but can also generate it independently, be it text, images, code or structured data. It is a sub-area of AI that focuses on transformer models, in particular large language models (LLMs) such as OpenAI's GPT, Google's Gemini or Meta's LLaMA.

 

These models are trained on billions of texts and use complex neural networks with billions of parameters to process language, structure and meaning. They generate new, contextually appropriate content based on an input prompt - be it text, program code, semantically structured data or even multimodal content such as images and speech. A core technical concept here is unsupervised learning or self-supervised learning: the model learns from large volumes of data without explicit labels and optimizes itself based on the probabilities with which words or tokens follow one another.

 

Our expertise lies not only in understanding these technologies, but also in the ability to integrate them into business processes in a targeted manner - whether for automation, decision support or new digital products. We also pay attention to topics such as data security, model control, governance and responsible AI.

Large Language Model (LLM)

These are AI models that are based on the Transformer architecture and are capable of understanding, processing and generating high-quality natural language. Well-known examples include GPT-4 (OpenAI), Claude (Anthropic), LLaMA (Meta) and Gemini (Google).
LLMs are trained on huge amounts of data and learn linguistic patterns, semantic relationships and contextual meanings. The result is models with billions of parameters that can provide very precise and human-like answers - even to complex questions.

 

Technical features

  • Transformer architecture: Basis for understanding word contexts across large text areas
  • Self-supervised learning: The model learns by predicting tokens in context - without manual labels
  • Generative behavior: The models create new text tokens for tokens based on the prompt
  • Fine-tuning: adaptation to specific tasks with minimal training effort or targeted tuning

 

We develop solutions with market-leading and open source LLMs and integrate them into the existing IT landscape on Azure OpenAI. Depending on the use case, we combine LLMs with:

 

  • Retrieval Augmented Generation (RAG)
  • Prompt Engineering & System Prompts
  • Custom embeddings & semantic search
  • Security, access control and auditability

 

With our expertise, we bring LLMs from the laboratory into reality - as scalable, secure and high-performance components in your business processes.

Retrieval Augmented Generation (RAG)

RAG is a hybrid approach in which an LLM is combined with an external knowledge store. The aim: the model can retrieve relevant information from documents, databases or APIs at each prompt before generating an answer.
Instead of relying purely on the internal (and limited) training knowledge of the LLM, the model is fed with up-to-date, company-specific or domain-specific information, which massively increases the accuracy, relevance and trustworthiness of the answers generated.

 

  • Embedding & Indexing: Unstructured content such as PDF, is converted into vector embeddings and stored in a vector store such as Weaviate, Pinecone, FAISS, Azure Cognitive Search.
  • Retrieval at runtime: A semantic search process is started at the prompt. The most relevant document snippets are retrieved and transferred to the LLM together with the user prompt.
  • Generation: The LLM now generates a response based on the prompt and the information retrieved - the result is more precise, more up-to-date and more comprehensible.

 

Advantages of RAG

  • Topicality: Content does not have to be included in the model training.
  • Cost efficiency: No fine-tuning necessary - instead, targeted querying of external sources.
    Explainability: Sources can be provided („grounding“).
  • Flexibility: Integration of any data sources - locally or in the cloud.

 

We develop RAG-based systems

  • Document and contract analysis
  • Enterprise Knowledge chatbots
  • Interactive assistants for customer service or HR
  • Compliance check tools with proof of source

 

We rely on proven components such as LangChain, Semantic Kernel or our own pipelines with open source LLMs - and consistently pay attention to data protection, scalability and relevance in the business context.

Azure OpenAI in the context of data protection

The Azure OpenAI Service combines the power of AI models with the highest standards of data protection, security and compliance - delivered via the Microsoft Azure infrastructure.

 

  • Data remains in the Azure region of its choice; for customers in Switzerland, use is possible via Azure regions in Switzerland North or Switzerland West. Data processing thus takes place exclusively within Switzerland or the EU, if desired
  • No model propagation - no training data All prompts and outputs (responses) are not used to train the models. The data never leaves the own instance and is not stored or analyzed by OpenAI itself
  • Microsoft Azure fulfills all relevant compliance requirements, including:
    • GDPR (EU & CH compliant)
    • ISO/IEC 27001, 27017, 27018
    • SOC 1, SOC 2, SOC 3
    • CSA STAR, ENS High

Azure OpenAI can be integrated into existing Azure services - including Azure Key Vault, Azure Private Endpoints, Virtual Networks and Role-Based Access Control (RBAC). This means that all communication remains encrypted, traceable and accessible in a controlled manner.

Explainability (XAI)

These are methods and approaches that make it possible to understand how an AI arrives at its answers. Transparency is particularly important for complex models such asLLMs or deep learning networks - especially in regulated industries such as finance, healthcare or public administration.

 

  • This creates trust: Users and decision-makers want to understand how a model works - especially if it supports business-critical decisions
  • Regulatory requirements often result in comprehensible decision-making processes
  • Transparency helps with debugging, quality control and avoiding bias
  • Those who can understand why an answer is given are more likely to accept it

 

Technical methods of XAI

  • Shows which input factors contributed most to the decision using SHAP or LIME
  • With LLMs, you can analyze which words the model paid particular attention to
  • Visualization of relevant areas in images or texts
  • Complex models are simulated by simplified, interpretable models
  • LLMs themselves can be used to justify their decisions in understandable language
  • Tables with the applied attributes can be displayed for the user so that they can see the source data directly

Reinforcement Learning (RL)

This is a method where an agent learns by trial and error to make the best possible decisions. The agent receives a reaction for each individual action. Good decisions are rewarded, bad decisions are punished. In this way, the system constantly evolves and develops a strategy for generating the best long-term benefit. 

 

The system is initially trained in a simulation environment until it goes live in a second phase. The system is also constantly evolving during production and can adapt to new conditions. 

 

Rewards and punishments are not arbitrary, but are defined in such a way that they encourage the desired behavior. The agent's goal is to maximize the sum of all rewards over many decisions - in other words, to find the best strategy in the long term.