We have all seen the news of large organizations putting their trust in a new GPT-powered chatbot on their website and having, well, not-so-ideal responses. Remember DPD, who had to ‘fire’ their new chatbot for insulting the company? Or Chevrolet, which inadvertently sold cars for $1 via its chatbot? These incidents highlight the challenges of innovating with GPT-powered chatbots. And addressing these challenges? Exactly what we at Oswald aimed to solve.
Introducing Oswald’s generative chatbots: a hybrid approach that allows you to control your chatbot’s responses and scope while taking advantage of the creativity large language models (LLMs) offer. By defining clear parameters for when the AI can autonomously respond and when it should defer to predetermined responses, you limit the potential for irrelevant or incorrect responses. Moreover, adding your own knowledge base in Oswald further aligns the AI-generated responses with your company’s information.
For the past eight years, we have dedicated ourselves to building the Oswald chatbot platform. The traditional way of building chatbots means defining clear intents and crafting specific responses to ensure the chatbots do exactly what you need, with no surprises. This meticulous approach, similar to systems like Siri and KBC Kate (at least - for now), has been our cornerstone, providing reliability and precision.
Initially, when ChatGPT was released, we wondered if traditional chatbots were becoming obsolete. A year later, we've seen clear examples of why they are not only still relevant but more important than ever. LLMs are incredibly powerful, but they can also make significant errors. Combining Oswald's traditional control with LLM creativity creates a new generation of bots you can confidently use on your website.
We’ve seen the strengths and limitations of both traditional and generative chatbots. Traditional ones offer unmatched control but require extensive manual input and can sometimes miss the nuances in user interactions. Generative chatbots, while capable of handling any question, can sometimes begin to hallucinate or are prone to prompt injections. So, what happens when you bring the two together? How does it work?
When developing your generative chatbot in Oswald, the first steps remain the same. It always starts with understanding the core purpose of your bot and identifying your target audience. You should begin by determining the specific tasks your chatbot should handle and considering the potential intents of your users. In addition, you should clearly define the scope of your chatbot’s capabilities — deciding what queries it should address and what’s beyond scope.
Once these basics are in place, here’s where the magic of our GPT integration changes the game. Traditionally, adding all intents and scenarios in your dialogue flow took days — and constant updates. Now, GPT’s integration significantly simplifies this process. You outline the areas where the LLM can autonomously generate responses, using your custom knowledge base as its guiding source. Yet, you retain the power to control specific, sensitive areas, ensuring your chatbot provides AI-generated responses only where it’s truly beneficial. And, to top it all off, you of course have all the features Oswald has been offering for years: a dedicated chatbot platform with human takeover, analytics, integration with other communication channels, testing, code responses, and more.
Considering costs, a standalone GPT bot may initially seem more cost-effective, but integrating Oswald with an LLM offers a far better long-term ROI. Using Oswald alone is possible but requires manpower to configure all conversations. Moreover, a traditional chatbot might not always grasp the nuance of a user’s questions, whereas an LLM can.
However, an LLM can also make mistakes, so integrating it into Oswald's controlled environment is interesting. For example, the integration guards against prompt injections — recall the Chevrolet example? — which can unexpectedly increase token usage (and thus, cost). Remember, each token generated incurs a fee.
Moreover, standard conversation flows, such as your chatbot’s initial greeting, are more efficiently managed in Oswald. When these responses are already predefined, you do not need to consume valuable tokens. Automating these recurring responses saves costs by reducing unnecessary token usage.
Another key advantage is the control you have over the chatbot's scope. Oswald ensures your chatbot remains within predefined parameters, avoiding the risk of generating irrelevant or off-topic responses, which can also lead to increased token usage.
The final aspect we want to highlight is security and privacy — a growing concern when processing confidential business and user information through LLMs. At Oswald, we prioritize our clients' and their users' security and privacy. For this reason, we have chosen to partner with Microsoft Azure for our OpenAI and AI Search services. Thanks to Microsoft Azure, we can provide a robust framework that is fully GPDR compliant, ensuring that your data remains protected and is not used by Microsoft to train its models. You'll have complete control over where these services are deployed, further enhancing data security and compliance with privacy regulations.
We're thrilled to introduce this new feature to the Oswald family. Let's put LLMs to work where they shine. Interested in seeing it in action? Request a demo, and we'll gladly show you — no strings attached.
