How overcoming off-the-Shelf Generative AI tools limitations to use it as an organisation?
Off-the-Shelf Generative AI tools might not be a good idea for your company…
Whether you’re looking for travel recommendations for your next trip to Tuscany, a summary of key characters in “To Kill a Mockingbird” or a concise explanation of quantum computing concepts, ChatGPT can be your ideal companion. What other conversational partner can boast omniscience, succinctness in responses, and endless patience? Of course, as everyone knows, you should exercise some discernment and not take all of its responses literally.
However, if you were considering automating your company’s customer service, neither ChatGPT nor its competitors would be of great help to you, because of the AI tools limitations. Worse, using an off-the-shelf generative AI could even harm your credibility or brand image. Trained in late 2021, ChatGPT 3.5 will have outdated knowledge of your product catalogue if it has been updated since then. Furthermore, its dominant tone may not accurately reflect the brand image you want to promote. Finally, from time to time, ChatGPT may start inventing products that do not exist in your catalogue or estimate the price of a service in a whimsical manner.
… Unless you customise it with contextual services and knowledge
The ideal solution to automate your customer service would be to combine the conversational abilities of ChatGPT (or any other Large Language Models (LLM)) with the factual accuracy that your customers expect. The tone of conversations should also be adjusted to establish a relationship in line with your company’s values. Similarly, the business vocabulary used should be precise to convey competence. Lastly, when a customer requests a price estimate, it must be accurate and, if possible, interpretable.
To achieve this, and to consider a multitude of options, the LLM should have access to both the latest version of the product catalogue and a price calculation module. It should also be able to draw on examples of reference conversations between one of your company’ subject matter expert and a panel of customers.
The general problem is the integration of data and services specific to an organisation into a general-purpose Language Model. Such systems are sometimes referred to as Augmented Language Models (ALMs) [1,2,3]. The motivations for this augmentation are numerous. Integrating external data can:
- Improve the factual accuracy of responses and, in particular, eliminate the hallucinations that LLM are known for
- Reduce the computational burden (= the number of parameters) of LLM by not requiring them to embed a list of factual knowledge into their parameters, which would be the result of a lengthy and expensive training phase
- Facilitate the immediate updating of knowledge external to the LLM
- Identify relevant sources to comply with regulations requiring precise references
- Adjust the style and vocabulary of a conversation or the output format of a response
- Protect confidentiality by entrusting all sensitive data to a private LLM
By integrating external tools
Indeed, integrating external tools can contribute to explain the logical reasoning that led to an answer by offering the possibility of examining the trace of the tools invoked to satisfy regulatory constraints on the traceability of processing. It also guarantees the accuracy of a calculation or estimate by delegating it to services with deterministic behaviour. And finally it enables external data integration!
The list of tools getting proportionally bigger as the subject develops, it can be easy to feel confused on where to start. You will find below a list of some tools that can enhance the reliability and reasoning capabilities of an LLM:
- Access to a knowledge base like Wikipedia or a relational database
- Arithmetic or symbolic computation modules (e.g., Mathematica)
- Logical inference engines (e.g., Prolog)
- Search engines to access current information (e.g., Google, Wikipedia)
- Interpreters for computer code (e.g., Python, JavaScript, or C++)
- Robot commands
- And, finally, any API!
