Retriever creation and activation for grounding
Key Takeaway
Saving a retriever doesn't make it usable. A prompt template can't use a retriever until you activate a version, and activating a new version automatically deactivates the version that is currently active.
A retriever searches a Data 360 search index and returns only the content that fits one use case. Prompt templates and agents use that content to ground LLM responses in accurate, current data. You build a retriever in AI Models (formerly Einstein Studio), activate a version, and then add the retriever to a prompt template in Prompt Builder.
Search indexes and retrievers
Teams often build a search index for general use, such as an entire knowledge base. A retriever applies that general search index to one use case, so it returns only the content the use case needs.
For example, imagine a search index built on an entire knowledge base. One retriever on that index could return only one category of articles. A second retriever could return only the last 90 days of content, and a third could return only content for one geographic region.
Prompt templates and agents use the content a retriever returns to ground LLM responses in accurate, current data, which is called Retrieval Augmented Generation (RAG). Retrievers can also serve flows and other features.