Skip to main content

Semantic Search Fundamentals

Vectara lets you build a semantic, LLM-powered search application. Semantic search is not just about finding data, but about understanding data and helping you answer questions about your data. This topic outlines what Vectara can do for this use case as well as why and how to employ these features for the best overall end-user experience.

Large Language Models (LLMs)​

LLMs are deep neural nets that are built with the task of specifically understanding human language. These models can be a great asset to many different use cases, including search and language generation.

These models generally work by reading immense amounts of text to build the model and then using that model to convert text into vectors, both at index and at query time. For many use cases, this obviates the need for many language rules of traditional keyword systems like synonym management, stemming, and phrase parsing because the LLM can inherently understand what the user is asking.

The team behind Vectara has built LLMs that work across a wide variety of languages and verticals. When you index data into Vectara or perform a search, also known as retrieval, the text is converted to one or more vectors via a LLM and then used to answer questions that your users have.

Zero-shot models​

Zero-shot models have an excellent understanding of language in general. They can understand and respond to the semantic meaning of questions without any additional tuning. This obviates much of the need for fine-tuning and specialized training on a particular dataset or in a particular vertical.

The Vectara platform makes extensive use of zero-shot models that have been developed by the team to allow your end users to query using the language and verbiage of their choosing and find the right documents, regardless of the domain your documents are in.

While zero-shot LLMs work very well in the vast majority of search use cases, there are some occasions where they struggle. In particular, many zero-shot LLMs don't work as well when users perform queries for things which have little semantic meaning.

For example, a UPC code, barcode number, or particular named configuration setting has little to no semantic meaning, and if you expect your users to perform this type of search, it's best to look into our hybrid search documentation to learn about how to blend neural search and keyword search.