noun_Email_707352 noun_917542_cc Map point Play Untitled Retweet Group 3 Fill 1

Improving response reliability for Large Language Models

How Retrieval-Augmented Generation can revolutionize the way we access information

Krzysztof Bączkiewicz / December 14, 2023

Retrieval-Augmented Generation (RAG) leverages the capabilities of Large Language Models (like GPT) while incorporating the capacity to search for and extract information from external sources. It addresses the reliability and accuracy issues faced by traditional language models by incorporating a search engine to fact-check and provide up-to-date information.

What is RAG

Retrieval-Augmented Generation, or RAG for short, is a cool new way to get the best of two worlds when it comes to finding information. It combines the power of large language models, like GPT, with the ability to search for and retrieve information from external sources. Large language models are great at generating text that sounds natural and makes sense, but sometimes they can get things wrong or make stuff up. That's where Retrieval-Augmented Generation comes in - it uses a search engine to find relevant documents that can provide accurate and up-to-date information. So, with Retrieval-Augmented Generation, you get the benefits of a large language model, but with the added bonus of being able to fact-check and provide reliable information.

LLMs' hallucination problem

LLMs without Retrieval-Augmented Generation have no way of checking the validity or the timeliness of the information they generate. They rely on their internal representations of the world, which are often biased or distorted by the data they are trained on. For example, an LLM might generate a text that is based on outdated statistics, or that reflects the opinions or prejudices of a certain group of authors.

LLMs are trained on static snapshots of text data, which do not capture the dynamic and evolving nature of the real world. For example, an LLM might generate a text that is based on a news article that was published several years ago, or that is no longer relevant or accurate. This can lead to misleading or outdated information that does not reflect the current state of affairs.

How RAG works

Retrieval-Augmented Generation is an awesome approach that combines the strengths of both LLMs and search engines to create texts that are more reliable and informative. RAG has three main parts: a query encoder, a retriever, and a generator. The query encoder turns your question into a search query. The retriever finds the most relevant documents based on the query and gives them to the generator. The generator then uses this information to generate a response to your question. It can either copy or rephrase the information, or use it as inspiration for generating new information.

To widen the capabilities of RAG approach search can be extended to include all kinds of data available for a use case. It is possible to search through relational databases, data lakes or even real-time data through available APIs. The idea is to ground LLM response on facts and reliable data only and to eliminate hallucination effect.

RAG's challenges

RAG is a really promising technique that can make the texts generated by LLMs even better in terms of quality and reliability. However, it also comes with some challenges that need to be tackled. Here are some of the main challenges:

Finding the right information. The retriever has to choose the most relevant data that can answer the user's question or prompt. But this is not always easy. The retriever has to deal with things like ambiguity, synonyms, polysemy, or anaphora, which can make it harder to understand and match the query and the data. The retriever also has to handle diversity, coverage, and redundancy, which can affect the amount and quality of information given to the generator.

Understanding what the user really wants. It's essential for the query encoder to fully grasp the user's requirements. This is especially important when the user's question or prompt is written in technical or domain-specific language, like medical, legal, or financial terms. The query encoder needs to overcome challenges such as specialized vocabulary, technical terms, and acronyms, to accurately represent and translate the query.

Finding reliable sources of information. It's not always a piece of cake to find sources that provide the most up-to-date and relevant information. Even when sources are available, it can be difficult to determine which one is the most reliable and trustworthy. This can result in the retrieval of incorrect or outdated information, which in turn can negatively affect the quality of the generated text.

We have done a lot of research and testing to tackle these challenges, and that's how AI Buddy came into existence.

Our proposal - AI Buddy

AI Buddy is a cutting-edge tool developed in Tietoevry Create, by a team of experts in building data-intensive solutions. It utilizes the power of Retrieval-Augmented Generation (RAG) to provide you with relevant and reliable information from various sources, based on your questions or prompts. AI Buddy can assist employees in accessing company knowledge by generating answers and solve problems from various internal sources, such as documents, databases, or any data via available API.
Find out more about AI Buddy

RAG real life use cases with AI Buddy

RAG-based solutions offer a more natural and engaging way to find, understand and discover information. Instead of just giving a list of links or snippets to read through or dashboards, these solutions generate answers that are customized to the user's question or prompt. They also provide an interactive and conversational experience by responding to the user's feedback or follow-up questions, eliminating the need to reformulate or refine the query.

Website search log analytics. Generative AI is used to deliver customer insights from search logs. Customer runs search for their products on multiple websites worldwide having traffic at the range millions queries per month. This is a great source of knowledge about what customers want and what they are looking for. LLM capabilities with RAG are being used to ask questions and have AI to analyze the logs delivering relevant answers.

Factory event viewer. The solution answers questions of what is going on in the factory using journal logs entries. If there is an issue or maintenance action taking place it is stored and documented. The trick is that this information is written in different local languages often using slang vocabulary. We are going to use generative AI to summarize and answer simple questions in English to support managers with the knowledge of what is happening in the factory without necessity of browsing through hundreds of logs.

Embracing challenges: unlocking the power of LLM services

Connecting data sources to LLM services is a task that requires attention to detail and adaptability. While initially seeming straightforward, it quickly becomes apparent that achieving the desired results involves overcoming various challenges.

Throughout this process, I have discovered the intricacies involved in tasks such as data preparation, prompt engineering, and fine-tuning the LLM. It became evident that making the search experience more effective often required adjustments to the search and ranking algorithms or even adapting the LLM to be more domain-specific.

Despite the challenges, witnessing the transformation of the data discovery experience has been both exciting and rewarding. Leveraging the power of language models has the potential to provide users with more accurate and contextually relevant results.

In conclusion, while connecting data sources to LLM services may not be without its difficulties, the benefits make the effort worthwhile. By embracing the challenges and refining the system, we can enhance data-intensive technologies and revolutionize the way information is accessed. Together, let's continue exploring the possibilities and unlock the full potential of LLM services.


Krzysztof Bączkiewicz
Lead Solution Consultant Tietoevry Create, Europe

Krzysztof is passionate about helping customers leverage their data to achieve their business goals. With expertise in information retrieval and business needs analysis, he delivers tailored solutions that drive growth and success.


Krzysztof Bączkiewicz

Lead Solution Consultant Tietoevry Create, Europe

Share on Facebook Tweet Share on LinkedIn