Published by Philip Creutzmann at


LLMs in Business Part 3: Unleashing the Power of Your Data

In today’s digital landscape, Large Language Models (LLMs) have emerged as a powerful tool for generating text, answering questions, and assisting users across various applications. However, the vast amount of data LLMs are trained on can sometimes raise concerns about the accuracy and reliability of their responses, particularly when trust in the source of information is paramount. Imagine a customer service chatbot providing answers based on unverified internet sources or a technical support bot that may not have access to the latest manuals and instruction documents. In such scenarios, trust in the source of information becomes crucial.

The Challenge: Trust and Reliability

LLMs are trained on large amounts of data from the internet, encompassing diverse sources and viewpoints. While this makes them versatile, it also introduces the challenge of ensuring that their responses are based on reliable, verified information. Working with out-of-the box LLMs imposes various challenges:

Hallucinations in LLMs: Hallucinations refer to the generation of text that appears plausible but is entirely fictional or factually incorrect. These hallucinations occur due to the statistical nature of LLMs, which are trained to predict the most likely next word or phrase based on patterns in their training data. Consequently, LLMs may generate content that aligns with their training data but lacks real-world accuracy, potentially leading to the spread of misinformation and undermining trust in their output.

Lack of Contextual Understanding: LLMs are incredibly talented, but they lack context about specific organizations, industries, or domains. This deficiency makes it difficult for them to provide accurate and context-aware responses. LLMs may generate generic or irrelevant information, leading to suboptimal results in tasks such as customer support.

Inconsistent Information: LLMs often produce inconsistent or contradictory responses due to the vast and diverse patterns in their training data. Users may find it challenging to trust LLM-generated information when faced with such inconsistencies, potentially leading to decision-making problems.

Privacy and Security Concerns: Fine-tuning LLMs using service providers may necessitate sharing sensitive or confidential information with the model. This raises legitimate privacy and security concerns, and organizations must exercise caution to avoid exposing proprietary data or compromising user privacy.

Limited Domain Expertise: LLMs are general-purpose models and do not possess specialized domain expertise. In industries or sectors that require specific knowledge or expertise, relying solely on LLMs may result in inaccurate or incomplete information.

The Solution: Retrieval Augmented Generation (RAG)

Enter the RAG architecture – a solution designed to tackle the trust and reliability issues associated with LLMs. The primary goal of this architecture is to limit the LLM’s responses to information from a predefined, trusted knowledge base, ensuring that it answers questions using only the provided information. The approach consists of two steps: First the relevant documents or databases are selected. Second, the context is provided to the LLM to answer the original questions. This approach not only empowers the system to offer answers but also supplies the source of information, effectively mitigating the risks associated with hallucinations and misinformation.

Application Examples: LLMs in Action

At our clients we implemented a wide range of use-cases. One of the big advantages is that the system provides not only the answer but also the source for further reading and validation.

  1. Customer Service Chatbot: A chatbot that draws its responses exclusively from FAQ documents created by the company. Customers can be confident that the information they receive is not only accurate but also comes from the organization itself, instilling trust and reliability in the interaction.

  2. Technical Support Bot: A technical support bot can access manuals, instruction documents, and troubleshooting guides as its exclusive knowledge base. This ensures that customers receive guidance based on the latest, verified information, reducing the risk of incorrect advice and enhancing trust.

  3. Corporate Document Search: Companies can use this architecture to create a powerful document search tool capable of finding and summarizing information from their internal documents, reports, and policies. Employees can trust that the results are drawn from authoritative sources within the organization.

  4. Natural Language Database Queries: Similar to documents whole database can be made accessible as a custom knowledge base. The LLM can translate natural language queries to SQL or similar database languages to retrieve and summarize the information. This can be used to generate custom reports and, when combined with other technologies, even creating dynamic charts on the fly.

Flexibility, Versatility, and Scalability

The custom knowledge base architecture offers remarkable flexibility, empowering organizations to tailor knowledge bases to specific needs while seamlessly integrating them into a wide range of applications. Unlike LLMs, these knowledge bases can be swiftly updated to stay current with evolving information, thus reducing the need for resource-intensive retraining. These systems can operate around the clock, utilizing any LLM service provider or on-premise LLM, and can be seamlessly integrated with other systems, such as those designed for permission limits.

Furthermore, they provide essential ethical safeguards by enabling organizations to implement controls that effectively mitigate the risk of biased or inappropriate content generation, thereby ensuring responsible AI use in today’s dynamic landscape.


In a world increasingly reliant on AI-powered interactions, trust in the source of information is paramount. The RAG architecture offers a powerful solution to ensure that LLMs provide responses grounded in trusted and known sources, overcoming the limitations of using LLMs in isolation. Whether it’s enhancing customer service, delivering accurate technical support, or enabling efficient corporate document searches, this architecture paves the way for more reliable and trustworthy AI-driven applications, making technology more reliable, useful, and responsible while maintaining the integrity of information and trust in the digital age.

Outlook and Resources

The frameworks around LLMs with custom knowledge bases are rapidly evolving and also the service providers are teasing solutions. By the time this blogpost is written, it is probably already outdated. A few frameworks to look out for are Langchain for seamless integration of various LLM services and prompt workflows. Combined with Llamaindex this enables the implementation of a custom knowledge base architecture.

Here are a few resources to explore:

  1. Generative AI exists because of the transformer – Excellent visual story telling about generative AI

  2. Langchain – For seamless integration of various LLM services, prompt workflows and RAG

  3. Llamaindex – Framework to index documents for search & retrieval

  4. OpenAI Blog – Introducing ChatGPT Enterprise

img img