Vectara raises $25M as it launches Mockingbird LLM for enterprise RAG applications

Vectara, an early pioneer in Retrieval Augmented Generation (RAG) technology, is raising a $25 million Series A funding round today as demand for its technologies continues to grow among enterprise users. Total funding to date for Vectara now stands at $53.5 million.

Vectara emerged from stealth in October 2022 and originally positioned its technology as a neural search as a service platform. It evolved its message to call the technology ‘grounded search‘ which the broader market now knows more commonly as RAG. The basics for grounded search and RAG is that responses to a large language model (LLM) are ‘grounded’ or referenced from an enterprise knowledge store, typically some form of vector capable database. The Vectara platform integrates multiple elements to enable a RAG pipeline, including the company’s Boomerang vector embedding engine.

Alongside the new funding, today the company announced its new Mockingbird LLM which is a purpose-built LLM for RAG. 

“We are releasing a new launch language model called Mockingbird which has been trained and fine tuned specifically to be more honest in how it comes up with conclusions and to stick to the facts as much as possible,” Amr Awadallah, co-founder and CEO of Vectara told VentureBeat in an exclusive interview. 

Enterprise RAG is about more than just having a vector database

As enterprise RAG interest and adoption has grown in the past year, there have been many entrants into the space. 

Many database technologies, including Oracle, PostgreSQL, DataStax, Neo4j and MongoDB to name a few all support vectors and RAG use cases. That increased availability of RAG technologies has dramatically increased the competition in the market. Awadallah emphasized his firm has numerous clear differentiators and the Vectara platform is more than just simply connecting a vector database to an LLM.

Awadallah noted that Vectara has developed a hallucination detection model that goes beyond basic RAG grounding to help improve accuracy. Vectara’s platform also provides explanations for the results and includes security features to protect against prompt attacks, which are important for regulated industries.

Another area where Vectara is looking to differentiate is with an integrated pipeline.  Rather than requiring customers to assemble different components like a vector database, retrieval model and generation model, Vectara offers an integrated RAG pipeline with all the necessary components.

“Our differentiation in a nutshell is very simple, we have the features required for regulated industries,” Awadallah said.

Don’t kill the Mockingbird, it’s the path to enterprise RAG powered agents

With the new Mockingbird LLM, Vectara is looking to further differentiate itself in the competitive market for enterprise RAG.

Awadallah noted that with many RAG approaches, a general purpose LLM such as OpenAI’s GPT-4 is used. Mockingbird in contrast is a fine-tuned LLM that has been optimized specifically for RAG workflows.

Among the benefits of the purpose built LLM is that it can further reduce the risk of hallucinations, as well as providing better citations.

“It makes sure that all of the references are included correctly,” Awadallah said. “To really have good extensibility you should be providing all of the possible citations that you can provide within the response and Mockingbird has been fine-tuned to do that.”

Going a step further, Vectara has designed Mockingbird to be optimized to generate structured output. That structured outputs could be in a format such as JSON which is becoming increasingly critical in enabling agent driven AI workflows.

“As you start to depend on a RAG pipeline to call API’s, you’re gonna call an API to execute an agentic AI kind of activity,” Awadallah said. “You really need that output to be structured in the form of an API call and this is what we support.”