Recently I chaired the Artificial Intelligence Accelerator Institute Conference in San Jose – in the heart of Silicon Valley. The event has brought together industry innovators from both large and small organizations, providing a wide range of perspectives. For example, the CEO of AI and ML testing startup of Kolena, Mohamed Elgendy and Srujana Kaddevarmuth, Senior Director, Data & ML Engineering, Customer Products, Walmart Global Tech discussed productization of AI solutions and ways to increase adoption. I especially liked the idea of a model catalogue from which data scientists can retrieve data sets and machine learning models that others have built rather than starting from scratch.
The presentations were excellent and ranged from higher level business problem and process approaches to deeper dives into the technology behind innovative approaches to solutions.
A presentation from Quantiphi discussed conversational AI – the use of chatbots to offload routine tasks from call center agents as well as provide “co-pilot” capabilities as it is listening in on discussions. Prabhpreet Bajaj, an associate practice lead at the company, walked through a section on ChatGPT and Large Language Models and pointed out the challenges that we have recently discussed in our webinar, writing and podcasts (link to ChatGPT article and webinar).
These include a wide range of issues, such as:
- the problem of “hallucinations” in which LLM’s create responses that sound plausible but are completely wrong,
- the challenges with exposing proprietary information to a public model,
- issues of source tracking and traceability,
- selecting and fine tuning of a foundational language model or a more specialized industry model, and
- developing a corporate ontology.
All these actions serve as preparation for training on a corporate or departmental knowledge base.
The need to point to an organization’s knowledge assets and not expose company IP is critical. The part that was glossed over a bit in presentations at the event as was the need for curated knowledge assets. When I asked “So, are you assuming that a company has a knowledge base that is correctly structured and organized?” The answer was “yes.” (See “Assume a Can Opener,” one of my prior articles, posted on our website). One cannot assume that what is needed as part of the solution is available.
We have to assume that in many cases, work needs to be done to structure, tag, and curate content. It is rarely can existing content be directly used by a ChatGPT type of application. Creation of a functional AI requires all of the things that we have preached over the years – correctly structured content models, taxonomies, metadata structures. These become embeddings in the data that LLM applications can retrieve.
The simple approach is for the LLM to process the query and use that processed query to retrieve information from a vector database. Content is ingested with “embeddings”- the correct metadata that adds context to the content such as product name or error code, etc., for an installation or troubleshooting guide, for example. The result is then processed and formatted by the LLM for conversational presentation to the customer.
The challenge right now is that every vendor has the talking points around how to use LLM’s, but many vendors are missing the nuances. Too often, they are glossing over a host of issues, such as the need for detailed knowledge process analysis, structured content development, metadata models that allow for contextualization of answers, the taxonomies that become part of the embeddings, and governance and metrics for tracking progress and ensuring continuous improvement.
Many of the initiatives using virtual assistants are tackling the lower hanging fruit of external customer service or reps who are supporting the customer experience. Few are tackling the overarching challenges of knowledge across the enterprise. Curating and managing content creation at the source, at the department and functional level, needs to be operationalized using centralized standards and processes, using component models so that content is chunked into semantically meaningful pieces.
As Large Language Models advance and become increasingly commoditized, the first adopters will have a progressively decreasing differentiator in the marketplace. The organizations that leverage all their knowledge assets are the ones that will build an increasing competitive advantage. Starting now will give the enterprise a leg up in our increasingly conversational world where the primary access to information systems will be through asking questions in natural language.