Off-the-shelf AI tools for adding chat functionality to packaging and processing equipment
A variety of web-based tools are cropping up that allow OEMs to offer AI-powered 24/7 chat support for CPG customers, relying on the OEM’s own documentation as source material.
Based on the OEM’s documentation, the AI can provide plain-language answers to questions. This tool from Scanmatics even cites and links to actual page numbers in the source documentation.
In past columns I’ve made the case for packaging and processing OEMs incorporate a “ChatGPT” style functionality into their machines. This can be an important bridge to address the skills gaps that their CPG customers are experiencing. My earlier columns have focused on embedding small language models directly on a PC in the machine, such as the HMI.
The argument for a direct embed is to allow access to real-time data while side-stepping the need for a persistent connection to the cloud, which is a non-starter for CPGs.
In this column we’ll explore the flip-side of this approach, using off-the-shelf tools that run in the cloud, but keeping them off the machine altogether. There are a few reasons why an OEM may want to do this vs. baking AI into the controls architecture directly.
1. Off-the-shelf AI tools are already built, slashing development and engineering time required for OEMs to learn about and incorporate AI models directly into their equipment. No learning Python, no downloading models and learning their intricacies, etc.
2. It leaves the sacred environment of the machine’s controls architecture intact and secure. No worrying about isolating the AI from the rest of the architecture.
3. Although small language models that can run on local machines do exist, they still require computing resources, and most require a PC with a GPU, which some industrial PC’s or HMI’s may have, but others may lack. Even on PC’s with a GPU, there are legitimate questions about the degradation in performance a locally running SLM could experience. There are also questions about the impact of taking computing resources away from the traditional HMI functionality or whatever else is running on that PC.
There are several off-the-shelf cloud-based tools that are worth checking out. All of these can run on any PC, tablet or phone, and are totally independent of the machine. These tools operate with what is known as Retrieval-Augmented Generation (RAG), a method where generative AI is directed to pull information exclusively from specific, user-provided documents rather than relying on pre-trained data or web knowledge. This approach greatly reduces the risk of hallucinations, as the AI generates answers based only on the trusted sources it’s given.
Wonderchat
First I’ll talk about Wonderchat, which is an application we are beta-testing here at PMMI. It powers the AI chat for PMMI ProSource, as well as the PACK EXPO show website, and we’re on the verge of rolling it out for for the PMMI website.
You can feed Wonderchat a variety of documents in formats including plain text, PDF, Microsoft Word, or PowerPoint. This is perfect for feeding in all kinds of written documents like manuals, though presumably it wouldn’t be able to make sense of pictorial information like wiring diagrams. (You can also upload numerical data in CSV format. Interestingly, you can upload an audio file or even a video file—up to 10 minutes in length max, though.)
Once it’s up and running, the Wonderchat chatbot can run on a website, such as a support website that you operate, or you can integrate it directly with a text messaging service. Customers can text their questions to a specific phone number and get an instant answer.
How do you make sure the chatbot returns accurate responses? It boils down to simply monitoring and testing. Wonderchat records literally every conversation. By having someone take the time to review every response, it is possible to flag a bad response and insert the correct answer so that the next time someone asks that question (or a similar one), it will give the right answer.
Another way to ensure accuracy is to not enable web search, so if the answer isn’t in your documentation, it won’t search the web and find something that may or may not be accurate. Wonderchat also has a way for you to specify key rules of what to say, and more importantly, what not to say. You can instruct Wonderchat as to the details of who your audience is, their background knowledge and technical level. In this way you can provide simpler answers for operators, and more technical answers for technicians, once it understands who it’s talking to. Finally, you can literally add instructions that tells the chatbot when it’s not 100% confident, how to handle its answer, which prevents hallucinations or inaccuracies.
When configuring your chatbot, switching underlying Large Language Models is as easy as choosing from a drop-down menu.In terms of cost and performance, there are over a dozen large language models that you as an OEM can select to power your AI chatbot. You will find that some work better than others. With testing and time you can zero in on the model that works best for you. Pricing varies from $19 to $500 per month, depending on usage. The model you select will impact usage (some are more resource intensive than others). Custom actions enable the chatbot to interact with external API’s.
Similar all-purpose off-the-shelf tools include Vectara and Querypal, though I haven’t looked at them in any detail.
Scanmatics
The other off-the-shelf tool I want to highlight is called Scanmatics. The company started out by providing a software tool to organize industrial automation documentation for manufacturing applications. OEMs of all stripes can organize their support documents in a private, web-based library for their manufacturing customers. Multiple manuals can be accommodated, whether it’s the OEM’s own manual, or manuals to critical machine components such as variable frequency drives. The tool was designed to eliminate the problem of frantically locating a manual after a machine goes down, sometimes years after the machine was installed and the original people involved have moved on.
Scanmatics core functionality consists of a library for OEM documentation that end users can access easily at any time.More recently, Scanmatics has added a chatbot on top of its core library functionality. This leverages the OEM’s existing documentation and basically eliminates the need for the OEM to do anything additional in order to provide chat support and functionality. CPG customers can ask technical questions and get technical answers, with links to the underlying support documents, right down to the specific page.
To accommodate the frequent changeover in personnel that CPGs experience, Scanmatics makes it easy for the CPG to install a durable QR code inside the machine cabinet (or wherever the CPGs want it). If an operator or technician is experiencing a problem with the machine, they just scan the QR code on the machine to pull up the relevant chatbot specific to the library of documentation for that machine.
(CPGs can also purchase the software and load the manuals themselves, allowing them to house documentation from multiple OEMs from one machine in one library, and giving that information a chat interface.)
Introducing AI chat support allows end users to ask a variety of questions, bring OEM documentation to life.Since Scanmatics is designed for industrial automation users, data and sessions are encrypted and protected through modern security techniques. No data from any OEM or CPG is used to train or inform the model for any other OEM or CPG. Like any RAG-based system, Scanmatics restricts its answers to what can be found in the OEM’s documentation, and does not refer to or access any information online. For CPG users, employees authenticate with a company email address, which means that new employees can access the tool without the OEM having to explicitly add them.
Though the company wouldn’t share pricing, it did say that OEM pricing for Scanmatics is based on how many CPG customers are using the functionality (based on number of active/operating machines in the field), and said that the pricing is designed to be affordable and reasonable for OEMs.
Although off-the-shelf tools mean OEMs give up some control over the AI experience, they offer a means for OEMs to start offering this functionality right away at a reasonable cost without consuming scarce engineering resources.
OEM Magazine is pleased to publish this semi-occasional column tracking the rapid advances in AI and how packaging and processing machine builders can leverage them to build next-generation equipment. Reach out to Dave at [email protected] and let him know what you think or what you’re working on when it comes to AI.
