Domain 2: Using OCI Generative AI Service (40%)

Domain 2 of the 1Z0-1127-25 Oracle Cloud Infrastructure 2025 Generative AI Professional exam covers the OCI Generative AI managed service end to end: pretrained models, dedicated AI clusters, fine-tuning, endpoints, inference APIs, security, and the Playground. At 40% of the exam (approximately 20 out of 50 questions), this is by far the heaviest domain. The exam syllabus identifies these topic areas:

1. Chat and embedding foundational models
2. Dedicated AI clusters (fine-tuning and hosting)
3. Fine-tuning base models with custom datasets
4. Model endpoints and deployment
5. Inference API parameters
6. Security architecture and IAM policies
7. OCI GenAI Playground

The exam format is 50 multiple-choice questions in 90 minutes with a passing score of 68%. Questions are scenario-based -- expect questions that test specific model IDs, parameter ranges, cluster unit types, and IAM resource names.

1. OCI Generative AI Service Fundamentals

OCI Generative AI is a fully managed Oracle Cloud service providing state-of-the-art large language models for chat, text generation, text embedding, and reranking. The service is accessed through the OCI Console (Playground), REST APIs, OCI CLI, and SDKs (Python, Java). (Overview)

Console path: Navigation Menu > Analytics & AI > AI Services > Generative AI

Two Operating Modes

On-demand mode caps response length at 4,000 tokens per run. Dedicated mode is uncapped up to the model's full context window. (Concepts)

Exam trap: On-demand text generation and summarization APIs are retired. Only chat and embedding are available on-demand. Generation/summarization models (e.g., cohere.command) can still run on dedicated clusters but not on-demand.

2. Pretrained Foundational Models

The service offers models from multiple providers. For exam purposes, the core models to know are the Cohere and Meta families. (Pretrained Models)

2.1 Chat Models

Cohere Command Family

Exam trap: The older cohere.command-r-16k model has a 16K context window, not 128K. Do not confuse it with cohere.command-r-08-2024 which has 128K.

Meta Llama Family

Exam trap: Llama 3.1 405B on-demand is only available in US Midwest (Chicago). All other regions require a dedicated cluster. The required cluster unit type is Large Generic 2 (not Large Generic 4, which was the older type).

Additional Providers (Newer Additions)

The service also offers Google Gemini (via Oracle Interconnect for Google Cloud, on-demand only), OpenAI gpt-oss models, and xAI Grok models. These are noted here for completeness but are less likely to be heavily tested since the exam syllabus was written around the Cohere/Meta core.

2.2 Embedding Models

All embedding models in OCI GenAI are Cohere Embed models. (Embed Models)

Key facts for the exam:

• Standard models output 1024 dimensions; Light models output 384 dimensions
• Maximum 96 inputs per run for text-only models
• Maximum 512 tokens per input for text-only Embed v3 models
• Text + image models support up to 128,000 tokens total across all inputs
• A 512x512 image consumes approximately 1,610 tokens
• Image input is API only -- not available in the Console Playground
• Embedding models cannot be fine-tuned

2.3 Rerank Model

Reranking is used in RAG pipelines to re-order retrieved documents by relevance before passing them to the LLM.

3. Dedicated AI Clusters

Dedicated AI clusters are single-tenant GPU compute resources for fine-tuning custom models or hosting endpoints. They are not shared with other tenancies. (Managing Dedicated AI Clusters)

3.1 Cluster Types

Exam trap: Fine-tuning clusters require significantly more GPU resources than hosting clusters. You cannot use a hosting cluster for fine-tuning.

3.2 GPU Unit Shapes

Unit shape names follow the format: <Instance Type>_<Number of Cards>. Examples: H100_X1 = H100 with 1 card. For A100 shapes, the memory size distinguishes variants: A100-80G vs A100-40G. The unit shape cannot be changed after cluster creation. (Creating Hosting Clusters)

3.3 Cluster Unit Types by Model

Each model requires a specific cluster unit type. These are critical for the exam:

Key facts:

• Maximum 50 endpoints per cluster (increase requestable)
• Multiple endpoints on the same cluster must use the same base model -- you cannot mix base models and custom models on one cluster
• Model replicas: Increase throughput by adding units (each replica = 1 additional unit)
• Cluster creation requires accepting commitment unit hours terms

3.4 Capacity and Scaling

• Default: 1 unit created per cluster
• Scale up by editing the cluster to add model replicas
• Each replica increases throughput proportionally
• Service limits control maximum units per shape (e.g., dedicated-unit-small-cohere-count, dedicated-unit-llama2-70-count)

Exam trap: To increase the Llama 3.1 405B hosting limit, you must request an increase of 4 units (not 1) because the multiplier is x4.

4. Fine-Tuning Base Models

Fine-tuning creates a custom model by training a copy of a pretrained base model on your own dataset. (Fine-Tune Models)

4.1 Fine-Tuning Methods

OCI GenAI supports two fine-tuning methods. The system automatically selects the method based on the chosen base model -- you do not manually choose. (Selecting a Fine-Tuning Method)

Exam trap: cohere.command-r-08-2024 supports both T-Few and LoRA. The Llama models support only LoRA. cohere.command-r-plus-08-2024 does not support fine-tuning at all.

4.2 Training Dataset Requirements

(Training Data Requirements)

Exam trap: The dataset must have exactly two fields: prompt and completion. Any other field structure will fail. The minimum is 32 pairs -- not 10, not 100.

4.3 Fine-Tuning Workflow

1. Create the training dataset in JSONL format
2. Upload the dataset to an OCI Object Storage bucket
3. Create a fine-tuning dedicated AI cluster (select the base model)
4. Create a new custom model (or new version of existing model)
5. Create a hosting dedicated AI cluster
6. Create an endpoint for the custom model on the hosting cluster
7. Test in the Playground or call via API

4.4 Hyperparameters

LoRA Hyperparameters (Meta Llama Models)

(Fine-Tuning Hyperparameters)

T-Few Hyperparameters (Cohere Models)

Total training steps formula:

totalTrainingSteps = (totalTrainingEpochs * datasetSize) / trainingBatchSize

Exam trap: T-Few defaults to 1 epoch with learning rate 0.01. LoRA defaults to 3 epochs with learning rate 0.0002. These are very different -- know which is which. Also note T-Few batch size range is 8-32 (default 16) while LoRA is 8-16 (default 8).

4.5 Fine-Tuning vs. Prompt Engineering

5. Creating Model Endpoints

An endpoint makes a model available for inference. Every model (pretrained, custom, or imported) requires an endpoint on a dedicated cluster for dedicated mode. On-demand models do not require explicit endpoint creation. (Creating Endpoints)

5.1 Endpoint Types

Exam trap: Imported models support public endpoints only -- private endpoints are not available for imported models.

5.2 Endpoint Configuration

Key settings during endpoint creation:

• Compartment: Where the endpoint lives (recommended: same as model)
• Model selection: Choose pretrained, custom, or imported model
• Dedicated AI cluster: Select existing active cluster or create new one
• Networking: Public (default) or Private endpoint
• Guardrails: Content moderation, prompt injection defense, PII handling (pretrained and custom only)
• Name: Auto-generated as generativeaiendpoint<timestamp> if not specified

5.3 Guardrails

Guardrails are optional safety controls configurable at the endpoint level for pretrained and custom models. They are not available for imported models. (Guardrails)

Key facts:

• Guardrails are disabled by default -- must be explicitly enabled
• Enabled via the ApplyGuardrails API or during endpoint creation in Console
• PII detection returns specific fields: text, label, offset, length, score
• Content Moderation tested on RTPLX dataset (38+ languages)

6. Inference API

The OCI Generative AI Inference API provides three primary operations: Chat, EmbedText, and ApplyGuardrails.

6.1 Chat API Parameters

These parameters control model output behavior. Know the ranges and defaults for each. (Cohere Command R+ (08-2024), Meta Llama 3.1 (405B))

Exam trap: Top K default for Cohere is 0 (disabled). Top K default for Llama is -1 (consider all tokens). Both effectively consider all tokens, but the numeric defaults differ.

6.2 Cohere-Specific Parameters

• Contextual: Fewer constraints; allows profanity and explicit content (entertainment/academic)
• Strict: Avoids sensitive topics (corporate/customer-facing)
• Off: No safety filtering

6.3 Embedding API (EmbedText)

(EmbedTextDetails API)

Required parameters:

• inputs (List[str]): Texts to embed (max 512 tokens each for v3 text-only)
• compartment_id: Target compartment OCID
• serving_mode: On-demand or dedicated

Optional parameters:

Exam trap: The input_type parameter is critical for RAG. Use SEARCH_DOCUMENT when embedding documents for storage. Use SEARCH_QUERY when embedding the user's search query. Using the wrong type degrades retrieval quality.

6.4 Token Estimation

Approximately 4 characters per token. This is a rough estimate used consistently across OCI GenAI documentation.

6.5 On-Demand vs. Dedicated Response Limits

7. Security Architecture

7.1 IAM Policies

OCI GenAI uses standard OCI IAM for access control. Only the Administrators group has access by default; all other users require explicit policies. (IAM Policies)

Aggregate resource type: generative-ai-family (covers all 11 individual resource types)

Individual Resource Types

Permission Verbs (Cumulative)

Common policy examples:

-- Full access at tenancy level
allow group GenAI-Admins to manage generative-ai-family in tenancy

-- Compartment-scoped inference only
allow group GenAI-Users to use generative-ai-chat in compartment AI-Prod

-- Embedding inference only
allow group GenAI-Users to use generative-ai-text-embedding in compartment AI-Prod

-- Manage clusters and endpoints
allow group GenAI-Ops to manage generative-ai-dedicated-ai-cluster in compartment AI-Prod
allow group GenAI-Ops to manage generative-ai-endpoint in compartment AI-Prod

Exam trap: To use the chat API, the verb is use on resource generative-ai-chat, not manage. The use verb is sufficient for inference. manage is only needed for creating/deleting resources.

Fine-Tuning Data Access

Training datasets in Object Storage require separate policies:

-- Upload datasets
allow group GenAI-Admins to manage object-family in compartment Data-Bucket

-- Read datasets during model creation
allow group GenAI-Admins to use object-family in compartment Data-Bucket

If training data and custom models are in different compartments, the user creating the model needs use object-family in the compartment containing the bucket.

7.2 Network Security

Private Endpoints: Restrict model access to traffic from within a VCN. (Prerequisites for Private Endpoints)

Prerequisites:

1. Create a VCN in the tenancy
2. Create a private subnet in the VCN
3. IAM policies: manage generative-ai-private-endpoint and manage virtual-network-family

The private endpoint is deployed as a VNIC in the private subnet. You manage the subnet's security rules and can optionally add network security groups (NSGs).

Service Gateway: For private network access to OCI services without internet traversal, use a VCN service gateway to reach OCI Generative AI through the Oracle Services Network.

7.3 Data Privacy and Model Isolation

• Dedicated AI clusters are single-tenant -- your data and models are not shared
• Custom model training data stays in your Object Storage bucket
• Fine-tuned models are private to your tenancy
• Guardrails provide additional content and PII controls

8. OCI GenAI Playground

The Playground is the Console-based no-code interface for testing models. It supports three modes:

Playground Features

• Parameter tuning: Adjust temperature, top-p, top-k, penalties, max tokens in the UI
• Token display: Shows input and output token counts after each generation
• Code export: "View code" generates code in multiple languages (Python, Java, etc.) with authentication pre-configured
• Vision support: Upload .png or .jpg images (max 5 MB) for multimodal models
• Example prompts: Pre-built prompts for quick testing
• Embedding visualization: 2D projection of embedding vectors showing semantic similarity

Embedding Playground specifics:

• Maximum 96 inputs per run
• File upload: .txt files only, newline-separated entries
• Truncate parameter defaults to NONE (returns error if exceeded)
• Export embeddings as JSON

Exam trap: In the embedding Playground, the Truncate parameter resets to NONE every time you click Clear. You must re-set it before each run if you want truncation.

9. Regional Availability

Model availability varies significantly by region. Key patterns to know: (Models by Region)

Availability symbols in Oracle docs:

• Check mark: Available (on-demand and dedicated)
• Check mark + o: On-demand only
• Check mark + d: Dedicated AI clusters only
• Check mark + G: Available through Oracle Interconnect for Google Cloud only

Exam trap: Google Gemini and xAI Grok models make external calls -- they route through Google Cloud or xAI infrastructure respectively. If data sovereignty is a concern, these models may not be appropriate. Cohere and Meta models run on OCI infrastructure.

10. Exam Focus: Common Traps and Pitfalls

1. On-demand response cap: Always 4,000 tokens. Dedicated mode is uncapped. If a question asks about maximum response length, the answer depends on the deployment mode.

2. Fine-tuning model support: Only cohere.command-r-08-2024, meta.llama-3.3-70b-instruct, and meta.llama-3.1-70b-instruct support fine-tuning. Command R+ does not. The 405B model does not. Embedding models do not.

3. T-Few vs. LoRA: The system auto-selects. T-Few is for Cohere; LoRA is for Llama (and also available for Cohere Command R 08-2024). You do not manually pick during model creation.

4. JSONL format: The only accepted format for training data is JSONL with exactly {"prompt": "...", "completion": "..."}. Minimum 32 pairs. One dataset per model. Auto-split 80/20.

5. Cluster unit types: Each model has a specific required unit type. You cannot mix models requiring different unit types on the same cluster. Know Small Cohere V2, Large Cohere V2_2, Large Generic 2, Embed Cohere.

6. IAM resource names: generative-ai-family is the aggregate type. For chat inference, the resource is generative-ai-chat with the use verb. For managing clusters, you need manage on generative-ai-dedicated-ai-cluster.

7. Private endpoints: Only for pretrained and custom models. Imported models are public only. Requires VCN + private subnet + IAM policies for both GenAI and virtual-network-family.

8. Embedding input_type: SEARCH_DOCUMENT for indexing documents, SEARCH_QUERY for queries. Using the correct type optimizes retrieval. Also available: CLASSIFICATION, CLUSTERING, IMAGE.

9. Embed dimensions: Standard = 1024, Light = 384. Embed v4 supports configurable dimensions (256, 512, 1024, 1536).

10. Guardrails disabled by default: You must explicitly enable content moderation, prompt injection defense, and PII detection. They do not run automatically.

References

• OCI Generative AI Service Overview
• Generative AI Concepts
• Pretrained Foundational Models
• Chat Models
• Embedding Models
• Cohere Command R+ (08-2024)
• Cohere Command R (08-2024)
• Meta Llama 3.1 (405B)
• Cohere Embed English 3
• Managing Dedicated AI Clusters
• Creating Hosting Clusters
• Creating Fine-Tuning Clusters
• Fine-Tuning Models
• Selecting a Fine-Tuning Method
• Fine-Tuning Hyperparameters
• Training Data Requirements
• Creating Endpoints
• IAM Policies for Generative AI
• Guardrails
• Models by Region
• Prerequisites for Private Endpoints
• EmbedTextDetails API Reference
• 1Z0-1127-25 Exam Syllabus (DBExam)
• 1Z0-1127-25 Exam Page (Oracle University)