From the course: LLM Foundations: Vector Databases for Caching and Retrieval Augmented Generation (RAG)

LLMs as a knowledge source

From the course: LLM Foundations: Vector Databases for Caching and Retrieval Augmented Generation (RAG)

LLMs as a knowledge source

“

As LLMs have become popular over the last many months, they are used for multiple purposes. This is because of their wide range of capabilities. There are two main capabilities of LLMs. The first is the language capability. LLMs can understand text and that too in several languages. They can learn the semantics and meaning from the text for reasoning. They can also generate human sounding text as responses. They can translate text from one language to another. The second is the knowledge capability. Given that LLMs are trained on a wide corpus of data, they can also answer questions related to their training data. They can help with distilling knowledge from these training sources and provide concise answers. There is hence the initiative to use LLMs as a source of knowledge. But this approach has a few shortcomings. First, LLMs can only answer questions based on the data they are trained on. LLMs are usually trained on public data that is available on the Internet. They are only good at answering questions related to that data. The answers from the LLMs may not be current. Their cut-off date is usually the date on which their original training data sources are extracted. LLMs have a tendency to hallucinate. They sometimes provide make-believe answers that are not factually correct. Also, for enterprise use cases, LLMs cannot answer questions based on enterprise or confidential data where this data is not part of the training dataset. It is possible to build custom LLMs using organizational data only, but that can prove to be expensive to build. It is also expensive to keep the LLM updated with new data on a daily basis. This is where retrieval-augmented generation comes in.

Contents