Diverse AI Solutions: Beyond Chatbots
Artificial intelligence has become widely known in society since the emergence of ChatGPT at the latest and is on everyone's lips. From an IT consultant’s point of view, a rising number of requests are mainly related to chatbot solutions, but we do try to show them the variety of different AI solutions. As a matter of fact, not every use case requires a generative approach. However, you could combine different strategies in order to get the best out of it.
A greater challenge arises, for example, from more complex structures in a service task that are stored in many different documents and requests. For example, I may have terms and conditions, specific instructions, questions about specifications or pricing, special requirements, all of which may be in different languages. In most conventional cases, you first have to spend a few hours screening the documents and summarizing the valuable and relevant findings. Accordingly, scaling is also limited, both in terms of time and content.
The “traditional” approach of a Machine-Learning-based solution would be the creation of a multi classifier in order to tag the requests accordingly. This already works well in many situations, but depends on a well-prepared and well-trained (and maybe even pre-labeled) training data set. The paradigm of Zero Shot Learning tries to break with that as prior training is not considered as mandatory anymore. However, when it comes to usable solutions for a business client, you will prefer a solution approach that is not too generic, but fitting to your specific needs.
For the implementation of enterprise solutions, LLM alone therefore only offers limited benefits, as previous examples will prove that have not been consistently integrated. Despite of recent studies showing that the bigger LLM could be more than just a stochastic parrot, leading to the situation in which the evidence is mounting that there is a basic understanding in these models (see “Random Graph Theory”), in terms of enterprise-ready solutions, you cannot rely on the randomness in connecting the dots. Hence, you will need a combination with semantic knowledge and hard facts.
Especially in the field of language processing with NLP and the now popular generative models, the methodology of so-called "Retrieval Augmented Generation" (RAG) promises to enrich the strengths of Large Language Models with fact-based knowledge and therefore deliver accurate results. This means a contextually relevant transfer of information. Despite existing for almost 5 years now, RAG technology (along with graphs) is one of the recent hot topics as it includes relevant sources that could be added to the response to a query. Hence, custom information will be integrated into a LLM without requiring another tedious retraining.
Another concern is to be found in privacy & security, but also transparency. On the one hand, you would like to have a proper and clear evaluation of the quality of your output, but without revealing your company data. It is also necessary to clearly define what happens to your data and how it is processed and stored. This also requires a proper implementation of your RAG technology.
Introducing Squirro - Out-of-the-Box Features for Enterprises
There are a couple of platforms that might help you with solving all of these issues. Among the tools leveraging this technology, Squirro has emerged as a particularly notable application. Let's explore the multifaceted applications of RAG and take a closer look at how Squirro is making a difference.
Following out-of-the-box capabilities are provided by Squirro:
- Pre-Trained ML Models, ready to use
- Pre-Built Data Connectors for an easy data integration
- Pre-Built Chat and Analytics Dashboards
- An End-to-End Automation Workflow (also visualized)
- No-Code Configuration & Integration
- Full Platform Observability
- Enterprise Security & Access Control
- SaaS / Managed Service Options
From a technology perspective, Squirro is a preferred solution for RAG not only because of its ability to use company and external data at scale, but also because it respects enterprise grade security by using Access Control Lists, Encryption and ISO standards.
Generative AI generally plays its strength for an enhanced self-service as it offers users an accurate full-answer automation and enables AI conversations with website content. Responses are AI-formulated but with high quality, which is important in preventing escalation. Cost reduction is a vital part of the service automation process. All of these topics (and more) are covered by the Squirro platform with its easy-to-handle UI.
Once a project is created (either empty or by choosing from a template), one could select data sources, load and enrich data. The integration process is ensured by different connectors and with a few single clicks you will be able to create a proper connection. Possible data sources could be e.g. from web, social media, a Python SDK, but also data from Jira could be indexed. As Business Intelligence tools are included, too, we could use a bidirectional approach for using aggregated reporting data in a workflow as well.
The AI studio offers training and publishing of machine learning models in terms of tagging, classification, entity extraction or trend detections.
Another mentioned advantage of Squirro is the convenient structure of creating automated pipelines, which minimizes “forgetting” important data preprocessing steps.
As the output is created, results will also show relevant sources of how the response was processed. This overcomes LLM hallucination by providing the evidence. This built-in feature comes in very handy and is quickly adjustable.
One final feature I’d really like to point out is the concept of communities used in Squirro. We could specify the interests of the users, whether they shall relate to clients, products, countries or a custom entity. Hence, the content will be personalized and optimized for each user.
So what are possible applications for this solution and who is to be targeted? The audience is broader than you think. For once, you could enhance your search engines not only based on keyword matching, but also on the context and nuance of the user’s query. Content creation and summarization is a key element of these models. Let’s also not forget Market Intelligence for businesses as vast amounts of data is available to provide insights and forecasts, making sense of discovering market trends and customer feedback.
SquirroGPT: Next-Generation Service Management
But also Customer Support, Service Automation (think about support ticket systems and service desks!) are glad if an AI-based tool would help them classify their incoming work accordingly, as this lowers their effort of directing the question to the correspondent topic and person responsible. A detailed support by retrieving and synthesizing information is therefore possible on the fly, significantly improving user experience. As for SquirroGPT for Service Management, employees will have the ability to instantly find relevant information from a knowledge base using natural language.
Bigger companies with different departments would profit from an automated classification which reduces the resolution time (benchmarks of Squirro state over 95% accuracy in ticket attribution and reduction of resolution time by more than 30%).