Vector database startup Zilliz Inc. said today that its newly updated Zilliz Cloud offering could be just the tonic for artificial intelligence practitioners looking to avoid the pitfalls of hallucinations.

Large language models such as ChatGPT have captured the imagination recently for their uncanny ability to provide humanlike responses to almost any question on any kind of topic. Yet Zilliz notes that these models are still far from perfect, with one of their main weaknesses being that they will often fabricate answers in the absence of correct information. A phenomenon that’s known in the AI industry as “hallucination,” it can be quite harmful in certain situations, such as if the AI is being used to field customer service questions.

Zilliz believes it can prevent these hallucinations. It cites OpenAI LP, the creator of ChatGPT, as saying that these fabricated answers can be minimized by supplying LLMs with external sources of domain-specific data, and that’s where Zilliz Cloud can help.

Zilliz Cloud is a vector database based on the open-source Milvus project that’s used to support AI applications. A vector database architecture is considered to be superior for AI models, especially machine learning, because these models typically take unstructured data such as documents, videos and user behaviors and convert them into vectors, which are essentially complex series of numbers. So it’s often a matter of finding which vectors are nearest or most similar to others.

In order to sort through and rank large numbers of vectors, a specialist vector database is required. Traditional databases are designed for tables and documents, so they’re inefficient for machine learning. Zilliz Cloud is different, since it has the ability to transform and index millions of vectors dynamically to answer the queries that are commonly thrown at AI models.

Zilliz Cloud is used to power AI models for product recommendation engines, semantic text search, targeted advertising, risk control and fraud prevention, among other things. However, its vector-native architecture also makes it ideal for LLMs.

By using OpenAI plugins to connect to ChatGPT, Zilliz Cloud can provide the foundation of a ChatGPT/VectorDB/Prompts-as-Code technology stack that enables LLMs to scale up their knowledge immensely, the company says, by accessing numerous additional data sources. Armed with more extensive knowledge, hallucinations should become a lot less common.

Chief Executive Charles Xie said that if AI is to fulfill its potential, it needs to become more trustworthy than it is now. “Hallucinations or wrong answers erode that trust,” he said. “With the billion-vector performance of Zilliz, we can help address that by expanding context and data retrieval.”

Image: DALL-E