The Copliot use case of a GenAI assis- tant, articulated by Patel, is particularly apropos to KM. According to Steve Gu, CEO of bitHuman, which offers interactive GenAI assistants that can mimic an actual person’s likeness, manners, voice, and image, “Employee training is actually a big area where corporate knowledge may be distilled and agents can onboard and train employees.” Implicit to these and other applications of virtual assistants is fine-tuning their generative models. “GenAI has this capability of trying out inferences from the contextual knowledge, the domain understanding, and driving knowledge graphs,” added Kavitha Chennupati, SS&C Blue Prism senior director of product management.
Traceability of outputs
Implicit to monitoring the results of generative models to minimize hallucinations and inaccuracies is the necessity of tracing their outputs to their sources. For textual applications of vector search, question-answering, summarization, and next-best actions, credible offerings are able to surface actual snippets of the textual data sources from which responses were derived. Galal mentioned that for this pervasive use case, “We do a citation. We found this, in this specific document, from here on page X, and this specific paragraph number.” Such detailed citations make it easy to trace model responses, validate them, and minimize hallucinations and erroneous outputs.
No matter which mechanisms a particular GenAI solution has in place, it’s incumbent upon organizations to continually observe the results, Patel said. “If we ask a question, how does it answer?” The benefit to KM practitioners is that users can ask more questions directly of the data sources to expedite information retrieval and accelerate work processes.
Knowledge curation
Another particularly cogent application of generative machine learning models for KM is the curation process for domain knowledge. Gu remarked that fine-tuning foundation models with such domain-specific knowledge is integral to eliciting relevant, correct responses. It involves “a data acquisition phase where clients give us [FAQs] and corporate documents. We put this knowledge as part of our system so the agent can answer questions in real time.”
Generative models are regularly used to populate knowledge graphs, compose business rules, and devise or augment entire taxonomies and ontologies. Dremio developer advocate Alex Merced noted that another application of this technology for KM entails “a built-in wiki so when people curate their data, they can create all the documentation around the data so they can make it available in the right context.”