Data 360 retriever grounding in Prompt Builder
Key Takeaway
You don't create retrievers in Prompt Builder. You build and activate an individual or ensemble retriever in AI Models, and then you insert it into the prompt template as a resource. At run time, it returns relevant content from a Data 360 search index to the template.
Retriever grounding uses retrieval augmented generation (RAG) to pull relevant knowledge that is indexed in Data 360 into a prompt template at run time. With it, a prompt can use unstructured content, such as knowledge articles, emails and chat transcripts. This added knowledge improves the quality of the LLM's responses and makes them more accurate, current and relevant. In Prompt Builder, you add this grounding by inserting a retriever into the template as a resource.
Search indexes and retrievers
A retriever is the component that searches knowledge indexed in Data 360 and returns the relevant pieces to the prompt template. It searches a search index, not raw records.
Data 360 prepares content for retrieval in these steps:
- Loading
- Chunking
- Vectorizing
- Storing
A search index holds that chunked, vectorized data. It is defined in a data space and associated with a Data Model Object (DMO).
A search index and a retriever do different jobs. A search index is often general-purpose, while a retriever applies that index to one use case.
For example, a search index might cover an entire knowledge base. A retriever built on that index could return only one category of articles, only the last 90 days of content, or only content from one geographical region.