Knowledge grounding with a Data Library
Key Takeaway
To ground an agent in published Knowledge articles, give it the Answer Questions with Knowledge standard action and assign a Knowledge-based data library that is ready. Data 360 must be set up first. At runtime, the library's retriever adds the most relevant chunks from the search index to the prompt, instead of the LLM reading the whole knowledge base.
An Agentforce Data Library connects agents and other AI features to your trusted content, such as Knowledge articles or uploaded files. The LLM gets domain-specific context from that content, so it can respond more accurately. When you save a library, the Data 360 parts it needs are built for you: data streams, a search index and a retriever. At runtime, the Answer Questions with Knowledge action uses those parts to pull the most relevant passages into the prompt.
Grounding and data libraries
A data library is how an agent gets grounding from your own content. Data libraries work with unstructured or semi-structured data, such as documents, web content or large text inputs in fields. They turn that content into searchable information.
Each library uses one of these data sources:
- Knowledge articles
- Uploaded files
- A custom retriever, which connects to more types of Data 360 data sources, including ensemble retrievers. Custom retrievers can also let companion orgs use libraries and search indexes configured in the home org.
The data source choice is permanent. You can't change a library's data source after you choose it. One library also can't use multiple data sources at the same time.
You can create several libraries. However, each feature, such as an agent, can use only one library at a time.
Web search isn't supported for new data libraries. Existing web search libraries are read-only. Those libraries have only a retriever and no search index. To search the web in an agent, use the Search the Web standard action. To search the web in a prompt template, use the web retriever in Prompt Builder.