OCI Generative AI Professional (1Z0-1127-25) Flashcards¶
71 cards from this guide's fact sheet and notes. Click a question to reveal the answer.
To study with spaced repetition, download flashcards.csv and import it into Anki or Quizlet (comma separated, first row is the header). Back to the study guide.
Exam logistics¶
How long is the OCI Generative AI Professional (1Z0-1127-25) exam?
90 minutes
How many questions are on the OCI Generative AI Professional (1Z0-1127-25) exam?
60
What is the passing score for OCI Generative AI Professional (1Z0-1127-25)?
68%
What does the OCI Generative AI Professional (1Z0-1127-25) exam cost?
USD 245 list price; Oracle periodically runs free certification windows
How long is OCI Generative AI Professional (1Z0-1127-25) valid for?
Re-versioned annually, indicated by the year suffix in the exam code
How is the OCI Generative AI Professional (1Z0-1127-25) exam delivered?
Oracle CertView, online proctored
Llm fundamentals¶
Cosine similarity (llm fundamentals)
the angle between vectors, ignoring magnitude. The usual default for text
Dot product (llm fundamentals)
considers magnitude as well as direction
Euclidean distance (llm fundamentals)
straight-line distance
Zero-shot (llm fundamentals)
instruction only
Few-shot (llm fundamentals)
include worked examples, which is often the cheapest large quality gain
Chain-of-thought (llm fundamentals)
ask for step-by-step reasoning before the answer, which helps on multi-step problems
System prompt (llm fundamentals)
role, tone, and constraints applying to the whole conversation
Transformer (llm fundamentals)
the attention-based architecture underlying modern language models
Attention (llm fundamentals)
the mechanism computing the relevance of every token to every other token
Decoder-only model (llm fundamentals)
the generative architecture used by most chat models
Encoder-only model (llm fundamentals)
the architecture producing representations, used for embeddings and classification
Token (llm fundamentals)
the unit of text a model processes, the basis for cost and context limits
Context window (llm fundamentals)
the total token budget for a request, spanning prompt, history, retrieved text, and output
Embedding (llm fundamentals)
a vector representation of text where semantic similarity is geometric closeness
Temperature (llm fundamentals)
the decoding parameter scaling randomness in token selection
Top-k sampling (llm fundamentals)
restricting token selection to the k most probable candidates
Top-p sampling (llm fundamentals)
restricting token selection to the smallest set of tokens reaching cumulative probability p
Frequency penalty (llm fundamentals)
a decoding parameter discouraging tokens already used often
Presence penalty (llm fundamentals)
a decoding parameter discouraging tokens that have appeared at all
Greedy decoding (llm fundamentals)
always selecting the highest-probability token
Stop sequence (llm fundamentals)
a string that terminates generation when produced
Few-shot prompting (llm fundamentals)
including worked examples in the prompt to demonstrate the task
Chain-of-thought prompting (llm fundamentals)
requesting step-by-step reasoning before the final answer
T-Few (llm fundamentals)
the parameter-efficient fine-tuning method used by OCI Generative AI
LoRA (llm fundamentals)
low-rank adaptation, a parameter-efficient fine-tuning technique
Perplexity (llm fundamentals)
a measure of how well a model predicts held-out text, where lower is better
LLM-as-judge (llm fundamentals)
using a language model to evaluate the outputs of another model against criteria
Oci generative ai service¶
Customer data isolation (oci generative ai service)
data submitted to the service is not used to train Oracle's base models, and dedicated clusters provide infrastructure isolation
OCI Generative AI (oci generative ai service)
the managed service providing foundation model inference, embeddings, and fine-tuning
On-demand inference (oci generative ai service)
per-request access to base models, billed by tokens
Dedicated AI cluster (oci generative ai service)
reserved, isolated capacity required for fine-tuning and for hosting custom models
Fine-tuning cluster (oci generative ai service)
a dedicated cluster that runs a fine-tuning job
Hosting cluster (oci generative ai service)
a dedicated cluster that serves a model through an endpoint
Endpoint (oci generative ai service)
the addressable serving target for a model on a hosting cluster
Custom model (oci generative ai service)
a model fine-tuned on your data, which can only be served from a dedicated cluster
T-Few (oci generative ai service)
the parameter-efficient fine-tuning method offered by the service
OCI Generative AI Agents (oci generative ai service)
the managed retrieval-augmented agent service over enterprise data
Compartment (oci generative ai service)
the OCI construct providing logical isolation and an access boundary
Private endpoint (oci generative ai service)
a private network path to the service, keeping traffic off the public internet
Content moderation (oci generative ai service)
service controls filtering harmful input and output
Service limit (oci generative ai service)
the per-tenancy, per-region quota governing resource consumption
Building llm applications¶
Reranking (building llm applications)
retrieve a wider set, then rerank with a more precise model and keep the best few
Metadata filtering (building llm applications)
constrain by tenant, date, document type, or access rights in the query, not after retrieval
Query rewriting (building llm applications)
have the model rephrase the question into better search terms
Faithfulness (building llm applications)
is every claim supported by the retrieved context?
Relevance (building llm applications)
does the answer address the question?
Correctness (building llm applications)
against a reference answer where one exists
Where it runs (building llm applications)
OCI Container Instances, OKE, or Functions for the application layer, calling the Generative AI service for inference
Observability (building llm applications)
trace each request end to end, logging the query, retrieved chunk IDs, prompt token count, model, latency, and cost. Retrieved chunk IDs are the single most useful thing to log, because most quality problems are retrieval problems
Caching (building llm applications)
cache embeddings for unchanged documents, and cache responses for repeated identical queries
Guardrails (building llm applications)
content moderation on input and output, and validation of any structured output before it reaches a downstream system
Re-indexing (building llm applications)
schedule it on document change, or the index silently goes stale
RAG (building llm applications)
retrieval-augmented generation, retrieving relevant text into the prompt before generating
Chunking (building llm applications)
splitting source documents into retrievable units, the largest lever on RAG quality
Chunk overlap (building llm applications)
repeated text between adjacent chunks so boundary-straddling information is not lost
Chunk metadata (building llm applications)
the source, section, and access attributes carried with each chunk, enabling citation and filtering
Top-K retrieval (building llm applications)
returning the K most similar chunks for a query
Hybrid search (building llm applications)
combining semantic similarity with keyword matching to improve recall
AI Vector Search (building llm applications)
Oracle Autonomous Database's native vector column type and similarity search
Retriever (building llm applications)
the LangChain component returning relevant documents for a query
Text splitter (building llm applications)
the LangChain component implementing a chunking strategy
Output parser (building llm applications)
the component converting model text into validated structured data
Conversation memory (building llm applications)
the mechanism carrying prior turns into the current request, and a major cost driver
Recall@K (building llm applications)
the proportion of queries for which the correct source appeared in the top K results
LLM-as-judge (building llm applications)
using a model to score outputs against criteria, itself requiring validation