Knowledge RAG grounding in prompt templates
Key Takeaway
A retriever connects a prompt template to the knowledge indexed in Data 360. To ground a prompt template with RAG, you add a retriever to it as a resource.
Retrieval augmented generation (RAG) grounds a prompt template with knowledge from unstructured content indexed in Data 360. Examples of this content include knowledge articles, emails, and chat transcripts. Before the prompt goes to the LLM, a retriever searches that index and adds the most relevant passages to the prompt. The LLM can then answer from your accurate, up-to-date, proprietary content, and you don't have to retrain or fine-tune the model.
How RAG works
RAG works in three steps:
- Retrieve relevant information from a knowledge store.
- Augment the original prompt with that information.
- Generate a response.
Many LLMs were trained on static, public Internet content. RAG supplies facts the model never learned, such as your latest policies or product details.