Grounding supplies a language model with necessary context inside the prompt, giving generation a factual base instead of leaving it to invent information. section 12.2.5
Retrieval-augmented generation (RAG) is a common grounding technique: retrieve relevant context, then use it to augment the model’s response. Technically, retrieval can use any search algorithm or database, although industry usage often implies semantic search with a vector database. section 12.2.5
Grounding has practical limits:
- It can improve difficult capabilities when the model receives helpful tools or context, but prompt cleverness without follow-up is insufficient. section 12.2.5
- RAG works best for simple questions; similarity search may miss the entities needed to answer a multistep question. section 12.2.5
A multistep question
“What is Gal Gadot’s husband’s current job?” requires identifying her husband before finding his occupation. A similarity search may retrieve many articles about Gadot but none about Jaron Varsano, leaving RAG poorly equipped to answer the complete question. section 12.2.5
