Cheat sheetGCP-03

Prompting, Grounding & RAG

GCP GenAI Leader / Prompting, Grounding & RAG

Quality comes from technique: prompt first, then ground / RAG, and fine-tune only if a gap remains -- cheaper levers before retraining.

Prompt design
First leverClear instructions, context, examples, and constraints -- cheapest way to lift quality.
Grounding
Trusted sourcesBase answers on your data or verified references; enables citations and cuts hallucination.
RAG
Retrieve then answerFetch relevant docs via embeddings and insert into the prompt so the model answers from them.
Fine-tuning
Last resortRetrain on domain data only when prompting and grounding cannot close the gap.
⚑

Instruct grounded models to answer only from retrieved context and to say when they do not know, then re-index sources when content changes.

RAG winSupport bot cites current KB articles; re-index on edit -- no retraining, less stale.
Anti-patternFine-tuning a model on today's docs, then watching answers go stale next month.
Order: prompt -> ground/RAG -> fine-tune (only if needed).
RAG = embeddings + vector search retrieve context into the prompt.
Grounding reduces, not eliminates, hallucination; constrain to context.
promptinggroundingragfine-tuninghallucination
review in 6d