Browse Oracle Certification Guides

OCI 1Z0-1127-25 Cheat Sheet

March 28, 2026

OCI 1Z0-1127-25 cheat sheet for key facts, traps, service mappings, and final review.

On this page

Use this for last-mile review. Keep it open during mixed OCI Generative AI questions and pair it with the Resources when you want Oracle’s exact GenAI phrasing. Strong answers usually separate model fit, retrieval quality, data boundary, safety, and deployment operations before touching prompt wording.

Read every GenAI scenario in this order

  1. Decide whether the issue is model capability, prompt design, retrieval, data access, safety, or operations.
  2. Check whether the system needs generation, summarization, classification, extraction, embedding, or tool use.
  3. If external facts matter, inspect the RAG layer before changing the model.
  4. If enterprise data is involved, confirm access control, metadata filters, tenancy boundaries, and logging.
  5. Optimize cost and latency only after correctness and safety are measurable.

OCI answer sequence

Use this when the stem mixes ingress, async delivery, reliability, security, or operations.

    flowchart TD
	  S["Scenario"] --> I["Classify the interaction mode"]
	  I --> E["Pick API Gateway, Events, Notifications, Streaming, or Functions"]
	  E --> R["Check retry, idempotency, ordering, and dead-letter behavior"]
	  R --> S2["Check Vault, IAM, private exposure, logs, and auditability"]

Fast lane picker

If the question is mainly about… Start with… Usual winning move
bad answers from good-looking prompts retrieval and grounding quality prompt wording is rarely the first fix
irrelevant or wrong documents chunking, metadata, filters, and top-k retrieval boundary first
unsafe or hostile input prompt-injection defenses and permission limits treat retrieved content as untrusted
slow or expensive inference context length, candidate set, caching, and model fit control tokens before chasing bigger infra
evaluation retrieval, groundedness, and safety separately do not collapse quality into one vague score
model selection task type, modality, context need, and latency budget bigger model is not automatically better
private enterprise knowledge RAG plus access-aware retrieval avoid putting every problem into fine-tuning
action-taking assistant tool permissions and auditability generation and execution are different risks

Canonical RAG flow

    flowchart TD
	  Docs["Documents"] --> Chunk["Chunk and Clean"]
	  Chunk --> Embed["Embeddings"]
	  Embed --> Index["Vector Index"]
	  Query["Query"] --> Retrieve["Retrieve Top-K With Filters"]
	  Index --> Retrieve
	  Retrieve --> Prompt["Prompt With Context"]
	  Prompt --> Model["Model"]
	  Model --> Answer["Answer With Citations or Grounding"]

Fast rule: better prompts rarely fix a broken retrieval layer.

Service-fit chooser

Requirement Stronger first fit
generate or summarize text with hosted foundation models OCI Generative AI style inference path
use enterprise documents without retraining RAG with vector search and metadata filters
build and manage custom ML workflows OCI Data Science style workflow
store and search embeddings vector index / vector database capability
call model output from an app endpoint, API, or application integration layer
automate actions from model reasoning tool/function integration with strict permissions

Model-choice reminders

Factor Better reading
context window limits how much source material and conversation state can be used
latency smaller or specialized model may beat a larger general model
cost token volume, model choice, retrieval volume, and reruns all matter
modality text, image, code, or embedding task shape should drive model choice
governance data handling and regional availability may eliminate otherwise attractive choices

Prompting reminders

Prompt need Stronger pattern
consistent answer format explicit output schema or structured instruction
better reasoning discipline specify role, constraints, and success criteria
reduce hallucination require use of provided context and say what to do when context is missing
extract facts provide fields, allowed values, and examples
unsafe request handling instruct refusal or escalation path, but do not rely on prompt alone

Retrieval chooser

Concern Strongest first lane Why
low relevance chunking and embeddings meaning may be represented badly
wrong tenant or version metadata filters boundary and freshness problem
too many weak documents top-k and ranking discipline context quality beats context quantity
fluent but unsupported output grounding quality and evaluation generation depends on retrieval quality
stale answer document freshness and index refresh path old chunks can beat a good prompt
mixed customer data tenant-aware filtering and authorization retrieval boundary is security boundary

Chunking and retrieval traps

Decision Too small Too big
chunk size weak context low precision
overlap wasted cost continuity breaks
metadata missing filters wrong tenant, version, or policy scope

RAG design decisions

Decision What to optimize
document cleaning remove boilerplate, navigation, duplicate headers, and low-signal text
chunk boundaries preserve semantic units such as section, table, or policy clause
embedding model match language, domain, and retrieval task
top-k enough candidates for coverage, not so many that context gets noisy
metadata filters tenant, product, version, date, policy, region, or entitlement
reranking improve final context quality when first-pass vector search is too broad

Embedding and vector-search traps

Trap Better reading
embedding everything without cleanup poor input creates poor retrieval
using only vector similarity when strict filters are needed security and freshness need metadata controls
increasing top-k to fix every miss more context can lower precision and raise cost
treating embeddings as reversible storage embeddings are search representations, not document originals
forgetting index refresh new or corrected documents need indexing before they can help

Evaluation signals

Layer What to measure
retrieval relevance, hit rate, filter correctness
generation groundedness, correctness, citation quality
safety prompt-injection resilience, leakage, policy violations
operations latency, token use, and failure rate
business fit task completion, escalation rate, and user correction rate

Evaluation traps

Trap Better reading
grading the whole system with one fuzzy quality impression score retrieval, generation, and safety separately
blaming the model before checking retrieval retrieval weakness often causes “model” weakness
assuming citation presence proves correctness citations can still point to poor or irrelevant retrieval
using only happy-path questions include adversarial, stale, missing-context, and permission-boundary tests
testing once before launch GenAI quality needs ongoing evaluation after content and behavior drift

Prompt-injection and safety cues

Concern Strongest first control
hostile instructions inside retrieved content treat documents as untrusted input
dangerous tool use limit permissions and allow only required actions
suspicious queries log, review, and monitor abuse patterns
cross-tenant data exposure strict metadata and access boundaries

Safety and governance controls

Risk Better control
prompt injection isolate instructions from retrieved content and constrain tool use
data leakage authorization-aware retrieval, redaction, and logging discipline
toxic or disallowed output safety filters, policy checks, and escalation path
unapproved model behavior evaluation gates and deployment approval
compliance evidence audit logs, data lineage, and retention policy
unsafe automation human approval or narrow tool permissions for high-impact actions

Tool-use and agent traps

Trap Better reading
model can call any tool because it is smart tools need least privilege and allowlists
successful natural-language answer proves action is safe generation and execution require separate controls
retrieved document can override system policy retrieved content is untrusted input
no audit trail for tool calls enterprise agent design needs action evidence
agent used for a simple lookup use RAG or direct retrieval when staged reasoning is unnecessary

Cost and latency controls

Pressure Better first move
large token cost cap context length and top-k intentionally
retrieval overhead use filters to shrink the candidate set
repeated embedding cost cache and reuse embeddings and indexes where appropriate
high latency reduce unnecessary context and expensive paths first
unpredictable spend set usage limits, monitor token volume, and cap retries
slow answer from many calls reduce chain depth before scaling infrastructure

Deployment cues

If the question is mainly about… Strongest first lane
safe rollout staged deployment and rollback path
quality drift continuous evaluation and monitoring
enterprise safety policy, permission, and data-boundary enforcement
operational confidence observe latency, errors, and token spend

Production monitoring map

Signal Why it matters
latency user experience and timeout risk
token usage direct cost and context-size pressure
retrieval hit quality source of many answer-quality failures
grounding score or review result detects unsupported answers
safety violation rate shows policy and abuse pressure
escalation / correction rate practical business-quality signal
index freshness prevents stale or missing source behavior

Troubleshooting order

Symptom Check first
fluent but wrong answer retrieval relevance and context grounding
cites wrong source metadata filters, chunk boundaries, and reranking
refuses too often safety policy, prompt constraints, and missing context
leaks wrong tenant authorization filter and retrieval boundary
too slow context length, top-k, chain depth, model size
too expensive token volume, retries, embedding refresh, model choice
works in test but not prod permissions, region, endpoint config, and data freshness

Decision order that usually wins

  1. Classify the failure as retrieval, generation, safety, deployment, or cost.
  2. Fix source quality, access boundaries, and filters before increasing model size.
  3. Measure retrieval and generation separately.
  4. Add safety controls beyond prompt wording.
  5. Optimize token and latency costs only after the answer is grounded and safe.

Last 15-minute review

If you only keep one thing from each lane… Remember this
RAG retrieval quality shapes answer quality
chunking chunk size and metadata both affect correctness
safety retrieved content is untrusted input
evaluation score retrieval, generation, and safety separately
cost top-k and context length are major levers
model choice fit the task, not the hype
deployment monitor quality, safety, latency, and spend together

What strong 1Z0-1127-25 answers usually do

  • identify whether the issue is retrieval, generation, safety, or deployment first
  • improve grounding and filters before reaching for more prompt complexity
  • treat prompt injection and data-boundary control as core system design issues
  • balance answer quality with latency and token cost instead of optimizing only one dimension
  • keep tool use, retrieval, and generation in separate risk lanes
  • use evaluation evidence instead of vague “better response” impressions
Revised on Sunday, May 10, 2026
Study Plan
Sample Questions