The potential sounds limitless, with the good marketing about ChatGPT Builder, a product from the current LLM leader OpenAI, it’s easy to think that everything can be automated from mundane tasks to complete organization management, all this with your own AI. All it needs is a strong LLM model, some data and the right prompt priming of the model.
Before celebrating, let’s look at the key concepts more in details. Each of these concepts are important to understand before we start assembling them together in an AI system. It is even more important to understand them before starting to fine tune this system. To keep this digestible, we’re skipping the whole digital transformation planing, data security aspect and the testing, monitoring and fine-tuning parts that would be required in an enterprise context. We’ll focus on a private person wanting to optimize their insurance contracts using off-the-shelf components.
The (LL-)Model
As the interconnected, hyper dimensional bunch of neurons, the model is the brains of the operation with a given set of “a priori” language and domain knowledge to understand information structures such as text, images and voice and to be able to communicate with their human counterparts. A useful or meaningful model can be created and fine-tuned only with enormous quantity of data, resources, power and time and con often only be run on expensive hardware allowing for a lot of parallel processing. Some companies out there are trying to create smaller models to allow them to be run on your own laptop or hand-held devices, but this is probably not the specialized case that concerns your everyday business. The training process for models had been roughly the same for more than 50 years until the transformers triggered this new era of AI, making training models magnitudes cheaper to create and making it at all possible to create models that could deal with sufficiently big datasets.
You can use open source models and have them run on your own hardware, or rent someone else’s hardware to run those open or closed source models. In the case of a private insurance contract comparison, you’d want to make sure that your model has been trained to understand the legal framework of your country and legal context. Different vendors are using different data set, so make sure to try out difference models and vendors to find the best match for you.
Prompt Priming
Similar to the filming of a film, imagine your model to be an actor on stage. You are the public, but before the scene gets filmed, you are allowed to prime your model as if you were the director of the film. You want to tell your model who it is, what it is thriving for, how creative, precise, cool or academic it should be, and how it should receive questions and respond to them. By priming your requests, you help the model understand not only your questions, but your expectations that were implicit until now. This greatly increases the quality of the answer you get.
Some Examples
In the role of a marketing specialist for the product, which is targeted to young adults between 18 and 25 years old, use your creativity to […]
In the role of a university professor in biochemistry, answer the questions with your complete knowledge about epigenetics. Remain precise and concise in your answer, citing your sources when possible. […]
Updating the prompts and it’s priming part is easy. What’s important is to keep notes of the most successful ones, and to reuse them in similar context. For a private person, just keep a notepad close to you, within an organization, you can create a small knowledge base in a shared confluence page for example.
Retrieval Augmented Generation (RAG)
You now have a model, and you know how to prime your prompts efficiently to ensure the best results. The issue is that your model might be aware of your domain (let’s say personal insurances here) but it is not aware of your current contractual situation with the different insurance parties. You want to share more relevant information with your model. To do so, you’ll want to give it some PDFs, text files, Excel sheets, pictures, access to your own contract database, phone and mail exchanges between you and different insurance companies. Letting it access the internet to get also information about new possible contracts with new assurance partners would help it make a meaningful suggestion.
At this point, this is where you notice that the strongest fine-tuning you can do to best leverage AI on your organization is to focus on the data that can be shared with your model. This is probably the closes you’ll get to “my AI” in a meaningful manner.
More precise implementation which can be used for business are able to not only use your own content, but then reference to you which content was used to generate an answer. This traceability can be key in multiple fields, especially regulated ones. In the case of a simple private implementation to compare insurance contracts, take a look at the file sharing capability of the GPT Builder which will load your files and use their data to answer you.
Data Management Enabling Efficient RAG
As you see, leveraging AI becomes quickly a question of data management. Unless you are a startup with geniuses from the top universities trying to create and improve the models, your challenges to leverage models that you can buy or rent is to have a meaningful data strategy. OpenAI is making it easier than ever to “click around and try something”. Don’t wait, start evaluating this solution today to see if this could improve your efficiency!
My own AI, how?
Your own AI often means, the best AI model, with your own prompt priming and your own data.
Once you’re ready to integrate the goodness of a useful AI in your company, make sure to make a luzid plan that will allow you to stay ahead of the curve while leveraging top LLM with your own organization’s data.
