OCI 1Z0-1127-25 Glossary: Key Terms
March 31, 2026
OCI 1Z0-1127-25 glossary of embeddings, vector search, prompt flows, and governance terms.
Use this glossary to clean up high-confusion OCI generative-AI terms before you go back into mixed sets. On this exam, terminology mistakes usually hide a systems-thinking mistake.
High-value terms
| Term |
What it means here |
Why it matters on the exam |
| Embedding |
a numeric representation that captures semantic similarity |
weak embedding strategy often causes weak retrieval |
| Evaluation |
the process of measuring usefulness, correctness, safety, or task fit |
this exam expects evaluation by layer, not by vibe |
| Fine-tuning |
additional training that changes model behavior beyond prompt-time controls |
candidates often overuse it when grounding is enough |
| Grounding |
supplying relevant external context so output is anchored to source material |
grounding is a core tie-break against hallucination |
| Hallucination |
unsupported, incorrect, or fabricated output that may still look fluent |
fluency is not evidence |
| Inference |
using a trained model to produce output from new input |
many questions hinge on what happens at inference time |
| Prompt injection |
hostile or manipulative instructions that try to override system behavior |
retrieved documents can carry untrusted instructions |
| RAG-style flow |
retrieval before generation so output is supported by source material |
this is a frequent exam decision lane |
| Safety control |
a rule, filter, permission, or process that reduces harmful output or leakage |
safety is not just one blocking keyword list |
| Service wrapper |
the product surface around a model capability |
wrapper and underlying model behavior are not the same thing |
Common confusion pairs
| Pair |
Clean separation |
| Grounding vs fine-tuning |
grounding supplies context at inference time, fine-tuning changes model behavior through training |
| Inference vs training |
inference produces output, training changes or builds the model |
| Prompt improvement vs model improvement |
prompt improvement changes the request, model improvement changes the underlying system |
| Embedding vs generated answer |
an embedding is a semantic representation, a generated answer is output text or content |
| Model capability vs service wrapper |
the wrapper is the product surface, the capability is what the model can actually do |
| Safety control vs quality control |
safety blocks or constrains harm, quality control measures whether the answer is good |
| Retrieval error vs generation error |
retrieval error brings in bad context, generation error mishandles the available context |
Fast recall anchors
| If you see… |
Think… |
| wrong documents |
retrieval quality |
| fluent but unsupported answer |
grounding and evaluation |
| bad output from risky source content |
prompt injection or safety boundary |
| expensive or slow answer |
context size, model fit, and delivery path |
If three terms blur together
| Terms |
Short reset |
| grounding, retrieval, embedding |
retrieval finds candidates, embeddings help similarity, grounding anchors generation with the chosen context |
| prompt engineering, fine-tuning, model choice |
prompting changes the request, fine-tuning changes behavior, model choice changes base capability |
| safety, governance, evaluation |
safety reduces harmful behavior, governance constrains who can do what, evaluation checks how well the system performs |
| inference, deployment, monitoring |
inference is the model call, deployment is how the system is served, monitoring is how you watch it in operation |
Route misses well
| If you missed because… |
Go next |
| you mixed up layers in the pipeline |
FAQ |
| you need operational tie-breaks fast |
Cheat Sheet |
| you need a paced rebuild of the weak lane |
Study Plan |
| you need the official Oracle or OCI source |
Resources |
Revised on Monday, June 15, 2026