Domain 3 of 4 · Chapter 3 of 3

Grounding and RAG

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing Generative AI Leader premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

  • What grounding is, and why it matters
  • Three sources of truth, and how RAG uses them
  • Google's grounding offerings
  • Sampling parameters: steering how the model generates
  • Exam patterns: grounding and generation

Grounding data sources: where the authoritative answer lives

ConsiderationFirst-party enterprise dataThird-party dataWorld data
What it isYour own private content: docs, wikis, databases, product catalogsLicensed or partner datasets you bring in but do not ownPublic information on the open web
Example questionWhat is our refund policy?What does our licensed market dataset report for this sector?What is the current published price of this product?
Google grounding offeringPrebuilt RAG with Vertex AI Search, or the RAG APIs for more controlThe RAG APIs (Vertex AI RAG Engine), pointed at your licensed datasetGrounding with Google Search
Best whenThe truth is internal and private to your businessThe truth sits in a specialized dataset you have licensedThe truth is public and changes over time

Decision tree

Must the answer come froma verified source of truth?No grounding neededtune sampling parametersWhere does theauthoritative answer live?Need customretrieval control?Grounding withGoogle SearchPrebuilt RAG(Vertex AI Search)RAG APIs(Vertex AI RAG Engine)NoYesOwn or licensed dataPublic webNoYes
Cheat sheet

Cheat sheet

  • Grounding answers from a source of truth, not the model's memory
  • Match grounding to first-party, third-party, or world data
  • RAG retrieves passages, then the model generates a cited answer
  • A RAG answer is only as good as its retrieval
  • Use prebuilt RAG with Vertex AI Search for fast cited answers over your data
  • Drop to the RAG APIs only when you need pipeline control
  • Use Grounding with Google Search when the answer is on the public web
  • Sampling parameters change how the model writes, not what it knows
  • Lower temperature for factual, repeatable output; raise it for variety
  • Top-p (nucleus sampling) widens or narrows the candidate token pool
  • Max output tokens caps length and cost, not quality
  • Safety settings filter harmful-content categories in the output
  • Skip grounding when the model already answers reliably
  • Grounding makes answers traceable through citations

Unlock with Premium — includes all practice exams and the complete study guide.

Also tested in

References

  1. Vertex AI grounding overview
  2. Vertex AI RAG overview
  3. Introduction to Vertex AI Search
  4. Adjust model parameter values