As companies explore generative AI more deeply, one of the more confounding issues is the hallucination problem, where if the model doesn’t know the answer, it simply makes one up, whether it makes sense or not. To work in business, finding a way to mitigate this problem is paramount.
Today, Writer, the generative AI writing platform, announced Knowledge Graph in public beta, a tool which enables companies to collect all of the current source content across an organization in a graph database. Company CEO May Habib says that by making use of the customer’s own content, Knowledge Graph can act as a kind of content validator, checking it against the generated content
“The great thing about an LLM is that if it has actual information about what you’re trying to do, it will use that information,” she said. “Knowledge Graph provides the ability to easily get the information that matters — databases, knowledge bases, shared drives, chats — into embeddings so that it can be used by the LLM in service of the very specific use cases our customers build around.”
It will automatically highlight content that needs to be fact-checked, while suggesting a replacement statement based on the most relevant content in the Knowledge Graph when appropriate. And Writer will show its work, displaying the source of the content where the replacement material came from, giving the human employee the ability to accept or reject the suggestion.
The company defines Knowledge Graph as “a structured representation of information stored in graph format, which allows for fast retrieval of specific pieces of information.” It can connect directly to the company’s content sources or customers can manually move content by dragging it dropping it into the Knowledge Graph interface.
It’s a separate repository, so it’s not actually part of the content systems that contribute to the graph. What’s more, companies can host the Knowledge Graph themselves if they wish, or they can let Writer manage it for them.
It’s by no means fool proof because nothing is, but it is a step towards trying to help companies reduce the impact of a known problem. Customers can choose how often they update Knowledge Graph, typically it’s weekly, but it could be more or less often if the customer wants to configure it that way, Habib said. The longer you take between updates, the more content that is not included in the accuracy check.
The company also announced the availability of the ability to self-host the Writer LLM, rather than having Writer do it. Some large organizations may want to more control over the content inside their organization and self hosting provides a higher level of control. The Knowledge Graph can connect to self-hosted LLM, as well.
Both the ability to self host and the Knowledge Graph are available in open beta starting today. The two products had been in a closed beta prior to this. Writer was founded in 2020, and has raised $26 million to date, per Crunchbase.
Writer introduces product that could help reduce hallucinated content in its LLMs by Ron Miller originally published on TechCrunch