Vectara Inc., a generative artificial intelligence search platform provider, today announced today that it has raised $28 million in seed funding led by Race Capital to empower developers with a new capability that greatly reduces AI errors when producing search results.

Vectara provides a cloud-based conversational generative large language model “search-as-a-service” that allows businesses to hold intelligent conversations with their own data, such as documents, knowledge bases and code. Its proprietary AI is similar to OpenAI LP’s ChatGPT but works on businesses’ own data. The company provides an application programming interface that gives developers quick access to its service to embed it into their apps, websites, chatbots and help desks.

The company released a new feature called “Grounded Generation” that augments queries with facts and data related to a company’s own data to reduce AI “hallucinations.” That’s what happens when an LLM, such as ChatGPT, confidently responds with a totally incorrect answer that’s insufficient, biased or completely random.

These issues have plagued the industry since the beginning, with prominent examples from Microsoft Corp.’s Bing Chat exhibiting bizarre behavior and Google LLC’s Bard making a factual error during its very first demo.

Industry LLM makers have attempted to handle these problems by increasing the amount of memory AIs have available to them, training datasets more tightly with the data they will use and constricting their answers, but hallucinations themselves still remain a concern.

With Grounded Generation, Vectara approaches the problem from a different direction. Instead of training the AI model with data from the company’s documents and knowledge base, it augments the conversational prompt for the AI with information pulled from the company’s data with search. This allows the responses from the AI to be constrained by valid responses based on the search query, which can then be fact-checked and provided with citations to the company’s documents and other data.

In addition to greatly reducing hallucinations, Vectara says, the effectiveness of search is also being improved by combining multiple search methods including semantic search, Boolean and exact keyword matching to provide highly relevant responses. The searches are capable of surfacing information even with misspellings, alternate meanings and different languages.

“These new features and capabilities make Vectara’s neural retrieval platform among the best in the world,” said Vectara co-founder and Chief Executive Amr Awadallah. “The breakthroughs our team has accomplished over the last eight months are changing the face of AI and how companies can safely use it to expand and improve their value propositions.”

Vectara can extract information from numerous document types including PDF and Office, as well as from code such as JSON, HTML, XML and CommonMark.

The company also stressed that because the LLM models are not trained on company data, such as indexed data or queries, customer privacy can be easily maintained. As a result, Vectara can provide businesses with customizable data retention models including allowing them to discard loaded documents after indexing and not retaining the original text or residual data.

In addition to the funding round the company established a new strategic board of advisors including members with expertise from tested veterans in the AI industry from companies and institutions including Cloudera, Stanford University and Northeastern University. Matei Zaharia, cofounder and chief technologist at Databricks Inc., will also join the newly founded board.

“Vectara’s new funding round will be instrumental in developing and expanding access to the platform’s breakthrough hybrid search and grounded generative AI features,” said Zaharia.

Image: Vectara