Stop AI Hallucinations: Complete RAG Architecture Explained with Enterprise AI Examples

Stop AI Hallucinations: How RAG Architecture Turns Raw Data into Precise Answers

A Complete Guide to Retrieval-Augmented Generation, Vector Databases, Embeddings, Enterprise AI, and Grounded LLM Responses

Generative AI | RAG | LLM | Oracle AI | Enterprise Architecture

1. Introduction: Why Traditional AI Needs RAG

Large Language Models (LLMs) have transformed how we interact with artificial intelligence. However, they have a fundamental limitation: their training knowledge does not automatically include current information or private enterprise documents.

When an LLM generates information without reliable supporting evidence, it may produce a confident but incorrect answer. This is known as an AI hallucination.

Retrieval-Augmented Generation (RAG) addresses this challenge by connecting an LLM to external knowledge repositories at query time.

Key Takeaway: RAG combines information retrieval with language generation, helping AI produce more relevant, verifiable, and source-grounded responses. RAG reduces hallucinations but does not guarantee their complete elimination.

2. What Is Retrieval-Augmented Generation?

RAG is an AI architecture that retrieves relevant information from external sources before generating an answer.

Instead of relying exclusively on training data, the model receives relevant documents as context.

Common enterprise knowledge sources include:

• PDF files and technical documentation
• Company policies and internal documents
• Standard Operating Procedures (SOPs)
• Web pages and product manuals
• Database administration runbooks
• Cloud architecture documentation
• Incident resolution knowledge articles

3. Complete Enterprise RAG Architecture Diagram

A RAG system consists of two interconnected processing pipelines:

• Offline ingestion: Prepares documents for efficient retrieval.
• Online retrieval and generation: Searches knowledge and generates answers.

Phase 1: Offline Data Ingestion

1. Enterprise Documents PDFs, Policies, Manuals, Web Pages

↓

2. Document Loading Extract Text and Metadata

↓

3. Text Chunking Split Documents into Smaller Segments

↓

4. Embedding Model Convert Text into Numerical Vectors

↓

5. Vector Database Index Embeddings, Chunks, and Metadata

Phase 2: Real-Time Query Processing

1. User Question Natural Language Query

↓

2. Query Embedding Convert Question into a Vector

↓

3. Retriever + Vector Search Search the Indexed Knowledge Base

↓

4. Relevant Text Chunks Retrieve Matching Passages and Sources

↓

5. Context + Question + Instructions Construct the Augmented Prompt

↓

6. Large Language Model Generate an Answer Using Context

↓

7. Grounded Answer Response with Supporting References

Architecture Note: The online retriever searches the vector database populated by the offline ingestion pipeline. Both phases share the same indexed knowledge base.

4. Step 1: Document Ingestion and Chunking

Enterprise knowledge is often stored in large, unstructured documents. Searching entire documents for every question can be inefficient.

Chunking divides documents into smaller text segments that can be independently indexed and retrieved.

Example: Oracle Database Documentation

Imagine a 200-page Oracle Database administration manual containing information about:

• RMAN backup and recovery
• Oracle Data Guard
• ASM storage management
• Database patching
• Performance tuning

If a DBA asks about RMAN recovery, the system can retrieve the relevant recovery sections instead of passing the entire manual to the LLM.

Good chunking strategies preserve context using document headings, overlapping segments, and source metadata.

5. Step 2: Embeddings and Semantic Understanding

An embedding model converts each text chunk into a high-dimensional numerical vector.

These vectors represent semantic characteristics of the text, enabling similarity-based retrieval.

Semantic Search Example

Question A: How can I restore an Oracle database?

Question B: What is the RMAN database recovery procedure?

Although the wording differs, both questions are closely related. An embedding model can help identify their semantic similarity.

6. Step 3: Vector Database

Generated embeddings are stored and indexed in a vector-capable database.

A vector database supports similarity searches to identify the most relevant stored chunks for a query.

Popular Vector Technologies

• Oracle AI Vector Search
• PostgreSQL with pgvector
• Pinecone
• Milvus
• Weaviate
• OpenSearch vector search

Vector retrieval may use cosine similarity, dot product, or Euclidean distance, depending on the embedding model and index.

7. Step 4: The Retriever

The retriever is the component that connects the user's question with the indexed knowledge base.

It performs the following operations:

1. Receives the user question.
2. Creates or obtains the query embedding.
3. Searches the vector database.
4. Identifies the most relevant document chunks.
5. Optionally applies filtering and reranking.
6. Returns selected passages and source metadata.

Example: Employee Leave Policy

User: "How many vacation days can I carry forward?"

Retriever: Finds the relevant section in the approved HR policy.

LLM: Generates an answer based on the retrieved policy, with a reference to the supporting document.

8. Step 5: Grounded LLM Generation

After retrieval, the application constructs an augmented prompt containing:

• The original user question
• Relevant retrieved text chunks
• Instructions for answer generation
• Source metadata for citations

User Question + Retrieved Context + Instructions
↓
Large Language Model
↓
Grounded Answer with Source References

The LLM uses the retrieved information to generate a coherent answer. Applications can instruct the model to acknowledge insufficient evidence rather than inventing unsupported details.

9. Traditional LLM vs. RAG Architecture

Feature	Traditional LLM	RAG-Based LLM
Knowledge	Primarily training knowledge	Training knowledge plus retrieved context
Private Documents	Not automatically available	Available through authorized retrieval
Updated Information	Requires external access or updates	Can retrieve recently indexed information
Hallucination Risk	Can produce unsupported answers	Can reduce risk through grounding
Source Attribution	Not inherently available	Can provide source citations
Model Retraining	May be needed for some knowledge changes	Often unnecessary for document updates

10. Real-World Use Case: Oracle DBA RAG Assistant

Consider an enterprise database team managing Oracle E-Business Suite, Oracle Database, AWS infrastructure, and disaster recovery.

The organization maintains hundreds of technical documents and operational procedures.

A RAG knowledge assistant can index:

• Oracle EBS ADOP patching procedures
• Oracle RMAN backup and restore runbooks
• Oracle Data Guard DR procedures
• Database cloning and refresh SOPs
• ASM storage troubleshooting guides
• AWS RDS operational documentation

Example DBA Question

"What is our approved procedure to convert Oracle Data Guard physical standby into snapshot standby?"

RAG Workflow:

1. The question is converted into an embedding.
2. The retriever searches approved DR runbooks.
3. Relevant snapshot standby sections are retrieved.
4. The LLM summarizes the documented procedure.
5. The answer provides references to the source runbook.

The DBA can verify the prerequisites, restore-point requirements, conversion steps, and rollback process before execution.

For production changes, RAG-generated guidance must still be validated against the environment and approved change-management procedures.

11. Enterprise RAG Best Practices

• Trusted Sources: Index approved and authoritative documents.
• Access Control: Enforce document-level permissions during retrieval.
• Freshness: Update indexes when source documents change.
• Chunking: Preserve context and document structure.
• Hybrid Search: Combine semantic and keyword retrieval where useful.
• Reranking: Improve relevance before sending context to the LLM.
• Citations: Include verifiable source references.
• Evaluation: Measure retrieval quality, faithfulness, latency, and cost.
• Security: Protect against prompt injection and unauthorized data exposure.
• Human Review: Require validation for high-impact decisions and production operations.

12. RAG and Agentic AI

RAG and Agentic AI are complementary architectures.

RAG retrieves knowledge, while AI agents may plan tasks, use tools, and execute authorized workflows.

For example, an AI-powered DBA assistant could retrieve a troubleshooting runbook, inspect read-only monitoring data, and recommend the next diagnostic steps.

Actions that modify production systems should require appropriate permissions and approvals.

13. Conclusion: From Raw Data to Grounded Intelligence

Retrieval-Augmented Generation provides a practical architecture for connecting generative AI with enterprise knowledge.

It separates knowledge storage from language generation, allowing organizations to update indexed information without continuously retraining the model.

Documents → Chunks → Embeddings → Vector Database → Retriever → Context + LLM → Grounded Answer

The goal of RAG is not simply to make AI sound more intelligent. It is to improve relevance, transparency, and factual grounding through reliable information retrieval.

Final Reflection

How could connecting AI models to your organization's internal documents, technical runbooks, and operational knowledge transform troubleshooting, automation, and decision-making?

Author: Punit Chitra

Oracle DBA | Cloud Database Technologies | Generative AI | Enterprise Automation

Punit Oracle DBA Blog

Topics: RAG Architecture | Generative AI | LLM | Vector Database | Oracle AI | Agentic AI

raditional AI vs Agentic AI vs Agentic RAG – Architecture, Differences & Oracle DBA Use Cases

Traditional AI vs Agentic AI vs Agentic RAG

Understanding the Next Evolution of Enterprise Artificial Intelligence: From Predictions to Autonomous Workflows and Context-Aware Reasoning

By Punit Kumar | October 28, 2026

Artificial Intelligence is rapidly transforming how organizations build applications, automate operations, and solve complex problems.

Traditional AI systems are often designed to make predictions, classifications, or recommendations. Modern agentic systems can go further by planning multi-step workflows, selecting tools, gathering evidence, and taking authorized actions.

Three approaches are particularly important: Traditional AI, Agentic AI, and Agentic RAG.

Quick Understanding

• Traditional AI: Predicts or classifies.
• Agentic AI: Plans, uses tools, and acts.
• Agentic RAG: Uses retrieval and reasoning to solve problems with relevant context.

1. What Is Traditional AI?

Traditional AI includes machine learning, predictive analytics, and rule-based systems designed to perform defined tasks.

Common applications include:

• Fraud detection and anomaly detection
• Customer behavior prediction
• Database capacity forecasting
• Predictive maintenance
• Workload classification

Traditional AI Architecture

Define Business Problem

↓

Collect Data

↓

Clean and Prepare Data

↓

Feature Engineering

↓

Train Machine Learning Model

↓

Evaluate Model

↓

Deploy Model

↓

Generate Predictions

↓

Monitor Model Performance

Oracle DBA Example: Tablespace Growth Forecasting

Consider an Oracle production database where a tablespace grows by approximately 100 GB every week.

A machine learning model can analyze historical storage usage and estimate when additional storage may be required.

Example AI Prediction:

The DATA tablespace may reach 90% utilization within the next 25 days.

The system provides a prediction. A DBA or a separate automation process typically determines the appropriate remediation.

Traditional AI does not necessarily require retraining from scratch when data changes. Some models support incremental learning, scheduled retraining, and automated updates.

2. What Is Agentic AI?

Agentic AI refers to AI systems that can pursue goals by planning, using external tools, evaluating results, and selecting subsequent steps within defined operational boundaries.

Unlike a single prediction or chatbot response, an AI agent can execute an iterative workflow.

Agentic AI Architecture

User Goal / Business Objective

↓

AI Agent / LLM

↓

Reasoning and Planning

↓

Select Tools and APIs

↓

Execute Authorized Operation

↓

Observe Results

↓

Evaluate and Decide Next Step

↓

Complete Task / Return Result

↺ If the goal is not achieved, replan and repeat within execution limits.

Oracle DBA Example: Intelligent Performance Investigation

Suppose an Oracle database experiences CPU utilization above 90%.

A conventional monitoring system generates an alert. An agentic AI workflow could perform the following:

1. Receive an Oracle Enterprise Manager or CloudWatch alert.
2. Collect active-session and SQL performance statistics.
3. Analyze wait events, AWR, ASH, and execution plans.
4. Identify likely causes of resource consumption.
5. Recommend remediation or initiate approved diagnostics.
6. Verify whether performance improves.

Example Agent Recommendation:

SQL_ID abc123 appears to be consuming significant CPU. Review its execution plan, workload changes, and optimizer statistics before considering remediation.

Advantages of Agentic AI

• Dynamic, multi-step problem solving
• Integration with external tools and APIs
• Automated investigation workflows
• Feedback-driven planning
• Reduced repetitive manual work

Agentic AI can adapt its workflow based on observations. However, this does not mean the underlying LLM automatically re-trains itself after every interaction.

3. What Is Retrieval-Augmented Generation (RAG)?

Retrieval-Augmented Generation, or RAG, combines a generative AI model with an external information retrieval system.

Instead of depending entirely on knowledge stored in model parameters, a RAG application retrieves relevant information before generating its response.

Basic RAG Workflow

User Question

↓

Query Processing

↓

Search Knowledge Sources

↓

Retrieve Relevant Information

↓

LLM + Retrieved Context

↓

Generate Grounded Response

Knowledge retrieval may use vector similarity search, keyword search, SQL, or hybrid search.

For Oracle environments, Oracle AI Vector Search can support semantic retrieval over embedded enterprise documents.

4. What Is Agentic RAG?

Agentic RAG extends conventional RAG by introducing agent-based planning, iterative retrieval, tool selection, and evidence evaluation.

Rather than retrieving documents once, the agent can decide whether additional searches are necessary and which sources should be consulted.

Agentic RAG Architecture

User Question / Objective

↓

AI Agent: Analyze and Plan

↓

Select Retrieval Sources

↓

Vector Search

SQL Database

APIs / Documents

↓

Retrieve and Combine Evidence

↓

Evaluate Relevance and Completeness

↓

Reason and Synthesize

↓

Verify Findings

↓

Evidence-Based Final Answer

↺ If evidence is insufficient, refine the query, search another source, and repeat.

Oracle DBA Example: ORA-12505 After Database Refresh

Imagine an Oracle E-Business Suite R12.2 environment encountering the following error after a database refresh:

ORA-12505: TNS:listener does not currently know of SID given in connect descriptor

An Agentic RAG system could:

1. Analyze the error and environment details.
2. Retrieve internal database refresh SOPs.
3. Search previous incidents involving listener errors.
4. Inspect authorized listener and database diagnostics.
5. Compare the current configuration with documented settings.
6. Recommend a validated correction.
7. Prepare a verified incident summary for knowledge-base review.

Example Recommendation:

The connect descriptor may reference an SID not registered with the listener. Verify the SID or SERVICE_NAME, database service registration, listener configuration, and Oracle EBS context settings.

This illustrates how Agentic RAG combines historical knowledge, live diagnostics, and iterative reasoning.

5. Traditional AI vs Agentic AI vs Agentic RAG

Capability	Traditional AI	Agentic AI	Agentic RAG
Primary purpose	Prediction and classification	Goal-oriented task execution	Context-aware reasoning and retrieval
Decision-making	Model-based or predefined	Dynamic planning	Dynamic, evidence-informed planning
External tools	Optional	Common	Common
Knowledge retrieval	Optional	Optional	Core capability
Multi-step workflows	Usually externally orchestrated	Yes	Yes
Autonomous actions	Not inherent	Possible	Possible
Memory	Model-dependent	Optional	Optional
Oracle DBA example	Predict storage growth	Investigate performance issues	Investigate using live data and runbooks

Important: Agentic RAG is a specialized application of Agentic AI, not a completely unrelated technology category.

6. Enterprise Use Cases for Oracle DBAs

Use Case 1: Database Performance Optimization

Traditional AI: Forecast CPU utilization.

Agentic AI: Investigate expensive SQL and recommend diagnostics.

Agentic RAG: Combine live performance metrics, historical AWR reports, and internal tuning procedures.

Use Case 2: Oracle EBS Patching

Traditional AI: Estimate patching duration.

Agentic AI: Coordinate prechecks and monitor approved workflows.

Agentic RAG: Consult Oracle patch documentation, ADOP runbooks, and historical patching failures.

Use Case 3: AWS RDS Monitoring

Traditional AI: Forecast storage consumption.

Agentic AI: Investigate CloudWatch alerts and initiate approved diagnostics.

Agentic RAG: Combine AWS metrics, architecture documents, and historical incidents.

Use Case 4: Oracle Data Guard DR Readiness

Traditional AI: Analyze replication-lag trends.

Agentic AI: Run authorized standby-health checks.

Agentic RAG: Combine transport/apply lag, DR runbooks, recovery objectives, and previous drill findings.

7. How to Build an Enterprise Agentic RAG System

A practical implementation should begin with a narrowly defined problem and expand only after reliability and security have been demonstrated.

1. Define the use case: Select a task such as database incident investigation.
2. Prepare knowledge: Collect SOPs, runbooks, incident reports, and documentation.
3. Implement retrieval: Use vector, keyword, SQL, or hybrid search.
4. Select the LLM: Evaluate accuracy, tool support, cost, and security.
5. Integrate tools: Connect approved database, monitoring, and cloud APIs.
6. Build orchestration: Implement planning, retries, execution limits, and approval checkpoints.
7. Evaluate: Test groundedness, accuracy, tool reliability, operational safety, and task completion.

Example Technology Stack

Component	Example Technologies
LLM	Enterprise-supported language models
Vector Search	Oracle AI Vector Search
Database	Oracle Database 23ai / 26ai
Orchestration	LangGraph, LangChain, custom workflows
Cloud	OCI, AWS
Monitoring	Oracle Enterprise Manager, CloudWatch
Integration	REST APIs, approved SQL tools
Security	IAM, RBAC, audit logging, approvals

8. Security and Governance

Autonomous systems require stronger safeguards when they can interact with production databases and infrastructure.

• Apply least-privilege access.
• Use read-only database access by default.
• Require approval for high-impact operations.
• Audit prompts, tool calls, and actions.
• Protect credentials and sensitive data.
• Defend against prompt injection in retrieved content.
• Implement execution timeouts and rate limits.
• Verify remediation results and maintain rollback plans.

Production DBA Security Rule:

Never provide an AI agent unrestricted SYSDBA access. Use scoped credentials, approved tools, and human authorization for operations that can affect availability, data integrity, or security.

9. Does Agentic AI Learn Automatically?

Agentic AI can improve future task execution through updated knowledge, retained task history, tool feedback, and refined workflows.

However, updating an agent's memory does not automatically retrain its underlying language model.

These are separate mechanisms:

• Memory: Retains useful context or history.
• Retrieval: Supplies relevant external knowledge.
• Planning: Adjusts steps based on observations.
• Training: Changes model parameters through a separate learning process.

10. Which AI Approach Should You Choose?

The best approach depends on the problem.

• Choose Traditional AI for well-defined forecasting, anomaly detection, and classification tasks.
• Choose Agentic AI for workflows involving multiple decisions, tools, and operational steps.
• Choose Agentic RAG when complex tasks require dynamic access to enterprise knowledge and verifiable evidence.

Recommended Enterprise Adoption Roadmap

Phase 1: AI Knowledge Search

↓

Phase 2: Agentic RAG Investigation

↓

Phase 3: Read-Only Diagnostics

↓

Phase 4: Human-Approved Remediation

↓

Phase 5: Controlled Automation at Scale

11. The Future of AI in Database Administration

Database administrators increasingly manage complex, distributed environments across on-premises infrastructure, Oracle E-Business Suite, cloud databases, and hybrid platforms.

Agentic AI and Agentic RAG can help DBA teams with:

• AI-assisted root cause analysis
• Database health checks
• SQL performance diagnostics
• Oracle EBS patching investigations
• Disaster recovery readiness assessments
• AWS database capacity and cost analysis
• Incident documentation and knowledge management
• Approved operational runbook automation

The objective is not to replace DBA expertise. It is to reduce repetitive investigation and help engineers focus on reliability, architecture, security, and critical decisions.

12. Final Thoughts

Traditional AI, Agentic AI, and Agentic RAG offer different ways to apply artificial intelligence in enterprise systems.

Traditional AI: Helps predict what may happen.

Agentic AI: Helps determine what to do and coordinate actions.

Agentic RAG: Helps make decisions using relevant, retrieved evidence.

The future of enterprise AI is not simply about larger language models. It is about combining reasoning, enterprise knowledge, tools, verification, and human oversight into reliable workflows.

For Oracle DBAs and cloud engineers, these technologies create new opportunities to transform reactive troubleshooting into intelligent, proactive operations.

The real question is no longer only: "Can AI answer this question?"

It is: "Can AI gather the right evidence, determine the next step, and safely help complete the task?"

Which Approach Are You Building Today?

Are you working with Traditional AI, Agentic AI, or Agentic RAG?

Share your thoughts and practical experiences in the comments.

About the Author

Punit Kumar is an Oracle Database and Oracle E-Business Suite technology professional specializing in enterprise database administration, AWS environments, high availability, disaster recovery, performance optimization, and automation.

This blog shares practical DBA guides, cloud technology insights, troubleshooting experiences, and emerging artificial intelligence concepts.

Blog: Punit Oracle DBA

Agentic AI Interview Glossary

The DBA's Guide to Oracle Agentic AI Foundations · Glossary

Every acronym, key idea and likely interview question from the series, explained in plain words with an Oracle DBA comparison. Bookmark it for revision before your exam or interview.

• Full forms
• Key concepts
• Interview Q&A
• How to answer

Full forms

The highlighted letters show where each acronym comes from. Learn the full form first; the meaning usually follows from it.

AI and model basics

AI

Artificial Intelligence

Machines doing tasks that need human-like thinking

ML

Machine Learning

Systems that learn patterns from data instead of fixed rules

GenAI

Generative AI

AI that creates new text, images or code

LLM

Large Language Model

The "brain" of an agent, trained on huge amounts of text

NLP

Natural Language Processing

Making computers understand human language

HITL

Human-in-the-Loop

A person approves each important step

Reasoning and retrieval

RAG

Retrieval-Augmented Generation

Fetch relevant documents first, then let the LLM answer from them

CoT

Chain-of-Thought

Reason step by step before answering

ReAct

Reason + Act

Thought → Action → Observation, in a loop

ToT

Tree-of-Thought

Explore several reasoning paths, keep the best

NL2SQL

Natural Language to SQL

Turn an English question into a SQL query

KB

Knowledge Base

The document library a RAG tool searches

Frameworks and protocols

MCP

Model Context Protocol

Open standard to plug any AI app into any tool ("USB-C of AI")

A2A

Agent2Agent protocol

Open standard for agents talking to other agents

SDK

Software Development Kit

Ready-made libraries for building apps, e.g. OpenAI Agents SDK

LCEL

LangChain Expression Language

Joins steps with a pipe: prompt | model | parser

API

Application Programming Interface

A defined way for software to talk to software

REST

Representational State Transfer

The common style for web APIs over HTTP

JSON

JavaScript Object Notation

The text format tools and LLMs exchange data in

JSON-RPC

JSON Remote Procedure Call

The message format MCP uses (version 2.0)

HTTP

Hypertext Transfer Protocol

Web transport; MCP's remote option is Streamable HTTP

SSE

Server-Sent Events

Older MCP remote transport, replaced by Streamable HTTP

stdio

Standard input/output

MCP transport for a local server on the same machine

OAuth

Open Authorization

Standard for secure delegated access to remote servers

Oracle and database

OCI

Oracle Cloud Infrastructure

Oracle's cloud

IAM

Identity and Access Management

Who can do what in OCI (policies, compartments)

ADB

Autonomous Database

Oracle's self-managing cloud database

HNSW

Hierarchical Navigable Small World

In-memory graph vector index: fastest, very accurate

IVF

Inverted File (Flat) index

Partition-based vector index: scales to huge data

ANN

Approximate Nearest Neighbor

Fast vector search using an index, slightly less than 100% accurate

KNN

k-Nearest Neighbors

Find the k closest vectors; exact search compares them all

ONNX

Open Neural Network Exchange

Model format you load into the database to create embeddings

VPD

Virtual Private Database

Row-level security that also applies to agent queries

PII

Personally Identifiable Information

Names, phone numbers, IDs: guardrails mask or block it

Key concepts

What each term means, and how to explain it using things a DBA already knows.

AI agent

An LLM that plans, uses tools and remembers, looping until a goal is met

The DBA wayThe on-call DBA who fixes the alert, not the one who only explains the error

Workflow (vs agent)

Fixed steps decided by code, with LLM calls inside

The DBA wayA scheduled RMAN script: same steps every time

Token

The small piece of text a model reads and writes (about ¾ of a word)

The DBA wayLike a block: the unit everything is counted and billed in

Context window

How many tokens the model can see at once; also its short-term memory

The DBA wayThe SGA: fast, limited, cleared when the session ends

System prompt

Fixed rules that set the agent's role and limits

The DBA wayThe runbook and change policy

Embedding

A list of numbers that captures the meaning of text

The DBA wayA fingerprint of meaning, not of spelling

Vector

The stored embedding; similar meanings sit close together

The DBA wayA row you search by closeness instead of equality

Semantic search

Finding results by meaning, not exact words

The DBA way"Tablespace full" also finds "no space in datafile"

Cosine similarity

Measures the angle between two vectors; the usual text metric

The DBA wayUsed inside VECTOR_DISTANCE(..., COSINE)

Chunking

Splitting documents into small pieces before embedding

The DBA wayLike splitting a large export into manageable files

Grounding

Tying answers to real data (tools, RAG)

The DBA wayAnswering from AWR, not from memory

Hallucination

A confident but made-up answer

The DBA wayA guess written into an RCA without proof

Function / tool calling

The LLM outputs a structured request; the app runs it

The DBA wayThe LLM writes the ticket; the operator executes it

Guardrails

Checks on input and output (injection, PII, toxicity)

The DBA wayChange control and audit for the agent

Prompt injection

Hidden instructions in input or tool results that try to hijack the agent

The DBA waySQL injection, but in plain language

Least privilege

Give tools only the access they need

The DBA wayA read-only DB user for a reporting agent

Handoff

One agent passes control to another

The DBA wayL1 hands the ticket to L2

Supervisor pattern

A manager agent delegates to specialists

The DBA wayThe incident manager in a P1 war room

Reflection

The agent critiques and improves its own output

The DBA wayReviewing your own RCA before sending it

Fine-tuning

Retraining a model on your data to change its behaviour or style

The DBA wayChanging the engine, not the fuel

RAG vs fine-tuning

RAG adds fresh facts at question time; fine-tuning changes how the model behaves

The DBA wayUse RAG for changing data like runbooks; it needs no retraining

Zero-shot / few-shot

Asking with no examples / with a few examples in the prompt

The DBA wayFew-shot = showing a junior two sample tickets first

Temperature

Controls randomness; low = consistent, high = creative

The DBA wayKeep it low for SQL and factual answers

Observability / tracing

A log of every step, tool call and result

The DBA wayThe audit trail and AWR for agents

Interview Q&A

The 15 questions you are most likely to hear. Each answer takes under 30 seconds to say.

Try answering out loud first, then tap a question to check yourself.

1What is an AI agent?

An AI agent is an LLM that can plan, use tools, remember and act in a loop until it reaches a goal. A chatbot answers once; an agent does the work, like an on-call DBA who checks, fixes and verifies.

2What are the components of an agent?

Five: the model, instructions, tools, memory and planning. I remember them as "My Instructions Tell Me Plans".

3Agent vs workflow: what's the difference?

In a workflow, code fixes the steps. In an agent, the LLM decides the next step based on what it finds. If the steps never change, a workflow is cheaper and more reliable.

4Does the LLM execute the tools itself?

No. The LLM returns a structured request with the tool name and arguments. The application runs it and sends the result back. That is function calling.

5Explain ReAct.

Reason + Act. The agent alternates Thought, Action and Observation until it can answer. It's how a DBA tunes a query: think, run the plan, read it, think again.

6What is RAG and why use it?

Retrieval-Augmented Generation. You retrieve relevant chunks from your documents and give them to the LLM with the question. It grounds answers in your data, reduces hallucination and needs no retraining.

7RAG or fine-tuning?

RAG for facts that change, like runbooks and policies. Fine-tuning to change the model's style or behaviour on a narrow task. Most enterprise use cases start with RAG.

8What is MCP?

Model Context Protocol, an open standard that connects AI apps to tools and data in one way. The host runs clients; each client connects to one server; servers expose tools, resources and prompts over JSON-RPC 2.0. Think of it as the Oracle listener for AI.

9Tools vs resources vs prompts in MCP?

Tools are actions the model decides to call. Resources are read-only context the application supplies. Prompts are templates the user picks.

10How do you build agents on OCI?

Create an agent with instructions, attach tools (RAG for documents, SQL for tables, custom functions for APIs), connect a knowledge base, and expose it through an endpoint with sessions and guardrails.

11What does Oracle AI Database add for agents?

It stores vectors next to relational and JSON data, so one query can combine semantic search and SQL. Existing security like privileges, VPD and auditing still applies, and there is no separate vector store to keep in sync.

12HNSW or IVF?

HNSW is an in-memory graph: fastest and very accurate, but needs the vector memory pool. IVF partitions vectors on disk and scales to very large data with less memory.

13What is Select AI?

Natural language to SQL inside the database, using DBMS_CLOUD_AI and an AI profile. Actions like showsql show the generated SQL, runsql runs it, and narrate answers in plain English.

14How do you make agents safe in production?

Input and output guardrails, least-privilege tools, human approval for irreversible actions, iteration and token limits, and full tracing. The same discipline as change control on a production database.

15Single agent or multi-agent?

Start with one agent. Move to multi-agent (supervisor, handoffs) only when tasks need clear specialisation, because each extra agent adds latency, cost and debugging effort.

How to answer any question

Use the same three steps every time. They show both knowledge and real experience.

Step 1

Define it with the full form

"RAG stands for Retrieval-Augmented Generation. It retrieves relevant documents and gives them to the LLM before it answers."

Step 2

Give the DBA analogy

"It's like answering a performance question from the AWR report instead of from memory."

Step 3

Connect it to Oracle

"In Oracle AI Database, the chunks sit in a VECTOR column, and I find them with VECTOR_DISTANCE using an HNSW index."

Don't know an answer? Say what you do know and how you would find out: "I haven't used that in production, but it works like X, and I'd check the OCI documentation for Y."

Part of The DBA's Guide to Oracle Agentic AI Foundations. For the full picture in one diagram, see the Agentic AI Cram Board. Independent study notes, not official Oracle material.

DBA AGENTIC AI FOUNDATIONS

The DBA's Guide to Oracle Agentic AI Foundations · 1Z0-1157-26

Agentic AI Cram Board

One picture that connects all six exam modules, a 60-second way to explain it, and a sticky note per module. Colours match: each box in the diagram is the same colour as its note.

40 questions60 minutes65% = 26 correct$0 on MyLearn

The whole picture

Follow a request from left to right. Every exam module sits somewhere on this path.

M1Agent basics M2Reasoning & memory M3LangChain & Agents SDK M4MCP M5OCI agents M6Oracle AI Database

Explain it in 60 seconds

If someone asks "what is agentic AI?", walk the diagram in this order.

1. A user's request reaches the agent endpoint. Sessions keep the conversation, and guardrails check what goes in and comes out.M5
2. Behind it sits the agent: an LLM with instructions, memory and a plan. It doesn't just answer; it works toward a goal.M1
3. It thinks in a loop: Thought → Action → Observation. That pattern is ReAct. Big jobs can be split across a supervisor and specialist agents.M2
4. To act, it calls tools. The LLM only asks; the application runs them. RAG for documents, SQL for tables, custom functions for APIs.M5
5. Outside tools plug in through MCP, a standard plug like an Oracle listener: client to server over JSON-RPC, exposing tools, resources and prompts.M4
6. The knowledge lives in Oracle AI Database as vectors, found by meaning with VECTOR_DISTANCE and HNSW or IVF indexes. Select AI turns English into SQL.M6
7. Developers build all of this with LangChain, LangGraph or the OpenAI Agents SDK.M3

Sticky notes

One per module. Read the "Say it" line out loud; if you can explain that sentence, you know the module.

Module 1

What is an agent?

• Agent = LLM + Instructions + Tools + Memory + Planning
• Trick: "My Instructions Tell Me Plans"
• Loop: Perceive → Reason → Act → Observe → repeat
• LLM requests a tool; the app executes it (function calling)
• Fixed steps = workflow · LLM picks steps = agent
• Human-in-the-loop → on-the-loop → autonomous
• Grounding (tools, RAG) cuts hallucination

Say itAn agent is an LLM that plans and uses tools in a loop until the goal is met.

Module 2

How agents think

• CoT: step-by-step reasoning, no tools
• ReAct: Thought → Action → Observation
• Plan-and-Execute: full plan first
• Reflection: critiques its own output
• Supervisor delegates · Handoff gives control away · Agent-as-tool keeps control
• Memory: short = context window · long = vector store
• Guardrails on input + output, least privilege, human approval, max iterations, tracing

Say itReAct is the DBA way: think, run a query, read the result, think again.

Module 3

The toolbox

• Loader → Splitter → Embeddings → Vector store → Retriever
• Prompt template · Chat model · Output parser
• LCEL: prompt | model | parser
• LangGraph: nodes, edges, state, checkpointer → loops and branches
• Agents SDK: Agent, Runner, @function_tool, Handoff, Guardrails (tripwire), Sessions, Tracing

Say itStraight line is a chain; loops and branches need LangGraph.

Module 4

MCP

• Open standard, "USB-C of AI", fixes the M×N problem
• Host holds clients; each client ↔ one server
• T-R-P: Tools = model · Resources = app (read-only) · Prompts = user
• Client offers: Sampling · Roots · Elicitation
• JSON-RPC 2.0 · stdio local · Streamable HTTP remote
• tools/list to discover, tools/call to run
• Consent, least privilege, OAuth

Say itMCP is the Oracle listener for AI: one standard way for any app to reach any tool.

Module 5

Agents on OCI

• Agent → Tools → Knowledge base → Endpoint
• Documents → RAG tool (with citations)
• Tables → SQL tool (give schema description)
• Actions → custom function tool
• KB stores: Object Storage · OpenSearch · Oracle AI DB
• New files? Rerun the ingestion job
• Endpoint: sessions, guardrails (moderation, injection, PII), citations, trace
• Access via IAM policies + compartments

Say itDocuments go to RAG, tables go to SQL, actions go to a custom tool.

Module 6

Oracle AI Database

• VECTOR(768, FLOAT32) column
• VECTOR_EMBEDDING + ONNX model in DB
• DBMS_VECTOR_CHAIN: text → chunks → embeddings
• ORDER BY VECTOR_DISTANCE(..., COSINE) FETCH FIRST 5
• HNSW = in-memory graph · IVF = partitions, scale
• Exact = 100% · APPROX ... TARGET ACCURACY = index
• RAG: load → chunk → embed → store → embed question → search → LLM → answer
• Select AI: DBMS_CLOUD_AI + AI profile; showsql shows, doesn't run

Say itThe database is the agent's memory: vectors next to tables, under the security we already manage.

Exam day

Before you click Start

• 40 questions · 60 minutes · ~90 seconds each
• Need 26 correct (65%)
• Answer easy ones first, mark hard ones, come back
• Cut two wrong options, then choose
• Match keywords: "plan first", "critic", "read-only", "in-memory"
• Never leave a question blank
• Quiet room, ID ready, other apps closed

Say itI know the picture, so I can work out any question from it.

Independent study notes for Oracle Agentic AI Foundations Associate (1Z0-1157-26). Check current exam details and service names on Oracle MyLearn.

The DBA’s Guide to Oracle Agentic AI Foundations (Oracle AI Database)

Agentic AI Foundations Series · Part 6 of 6

Part 6: Oracle AI Database – The Agent's Memory

1Z0-1157-26 exam prep for Oracle DBAs

Oracle AI Database stores vectors right next to your tables, JSON, and graphs. An agent can search by meaning and run SQL in one place, under the security you already manage. For a DBA, this is home ground.

What is a vector, really?

An embedding is a list of numbers that captures meaning. "Tablespace full" and "no space left in datafile" use different words but land close together. "Listener down" lands far away. Searching by closeness = semantic search.

The new SQL pieces

• VECTOR data type – a column for embeddings, e.g. VECTOR(768, FLOAT32).
• VECTOR_EMBEDDING() – creates embeddings inside the database using an ONNX model you load with DBMS_VECTOR.LOAD_ONNX_MODEL. You can also call external models such as OCI Generative AI.
• DBMS_VECTOR_CHAIN – the pipeline helpers: UTL_TO_TEXT (extract), UTL_TO_CHUNKS (split), UTL_TO_EMBEDDINGS (embed).
• VECTOR_DISTANCE() – measures closeness. Order by it and fetch the top rows:

SELECT doc_id, chunk_text
FROM   runbook_chunks
ORDER  BY VECTOR_DISTANCE(embedding, :q_vec, COSINE)
FETCH  FIRST 5 ROWS ONLY;

Distance metrics: COSINE (most common for text), EUCLIDEAN, DOT, MANHATTAN, and HAMMING/JACCARD for binary vectors.

Vector indexes: think B-tree vs bitmap

Just as you choose index types for workloads, you choose vector indexes:

• HNSW – an in-memory graph of neighbors. Fastest and very accurate. Needs the vector memory pool (VECTOR_MEMORY_SIZE).
• IVF – groups vectors into partitions on disk. Scales to huge data with less memory.

Exact search compares every vector: 100% accurate, slower. Approximate search uses the index: FETCH APPROX FIRST 5 ROWS ONLY WITH TARGET ACCURACY 90.

Exam hook: HNSW = memory + speed. IVF = partitions + scale.

RAG in eight steps

Be ready to put these in order:

1. Load documents
2. Chunk them
3. Embed the chunks
4. Store in a VECTOR column and index it
5. Embed the user's question
6. Similarity search for the top-k chunks
7. Send question + chunks to the LLM
8. Return a grounded answer with sources

Select AI: talk to your data in English

• Package DBMS_CLOUD_AI; you create an AI profile (provider, credential, model, allowed tables).
• Actions: runsql (default, runs it), showsql (shows SQL only), explainsql, narrate (answer in plain language), chat.
• Select AI with RAG adds vector search over your documents; Select AI Agent builds agents with tools inside the database.

Why keep the agent's memory in the database?

One source of truth, transactional consistency, and your existing privileges, VPD, and auditing apply to every agent query. No separate vector store to sync.

Remember for the exam

• VECTOR type, VECTOR_EMBEDDING, VECTOR_DISTANCE; COSINE is the usual metric.
• HNSW in memory, IVF partitioned; approximate = index, exact = full compare.
• Select AI = DBMS_CLOUD_AI + AI profile; showsql shows without running.

Self-check

1. Which vector index lives in memory? (HNSW)
2. Which Select AI action shows the generated SQL without running it? (showsql)
3. In RAG, what happens right after chunking? (Embedding the chunks)

That's the series. Take the free practice exam on Oracle MyLearn next.

The DBA’s Guide to Oracle Agentic AI Foundations (OCI agent service)

Agentic AI Foundations Series · Part 5 of 6

Part 5: Building Agents on OCI – Tools, Knowledge Bases, and Endpoints

1Z0-1157-26 exam prep for Oracle DBAs

OCI gives you a managed agent service: you describe the agent, attach tools and data, and publish it through an endpoint. No servers to patch. (Oracle renames services often, so match the names to what you see in the course videos.)

The help-desk analogy

Picture building an "Ask the DBA team" assistant for your company.

• The agent is the assistant, with instructions like "Answer database questions politely; never share passwords."
• Its tools are what it can do.
• Its knowledge base is the library of runbooks and SOPs it reads from.
• The endpoint is the front door your users' apps knock on.

The tools (know which one fits which question)

Need	Tool to choose
Answer from PDFs, runbooks, SOPs, with citations	RAG tool + knowledge base
Answer "How many open SRs this week?" from tables	SQL tool (natural language to SQL; give it a schema description)
Create a ticket, call a REST API, run your function	Custom / function-calling tool
Use another service or agent	API endpoint / agent tool

Exam hook: Documents → RAG. Tables → SQL. Actions → custom function.

Knowledge bases: where the agent reads

A knowledge base points to a data store:

• Object Storage – drop files in a bucket; OCI chunks, embeds, and indexes them for you (service-managed).
• OCI Search with OpenSearch – bring an existing search index.
• Oracle AI Database vector store – bring your own vectors, with your own database security.

Files get in through a data source and an ingestion job. If you add new runbooks, rerun ingestion.

The endpoint: front door settings

• Sessions – keep conversation context across turns (with an idle timeout).
• Guardrails – content moderation, prompt-injection protection, PII protection.
• Citations and trace – show sources and the agent's steps.

The OCI pieces around it

• OCI Generative AI – the hosted chat and embedding models the agent uses; dedicated AI clusters for fine-tuning or private hosting.
• IAM policies and compartments – who can create agents and read the data. Same idea as granting privileges.
• Object Storage – usually where source documents live.

Remember for the exam

• RAG tool for documents, SQL tool for tables, custom tool for APIs.
• Knowledge base stores: Object Storage, OpenSearch, Oracle AI Database.
• Turn on sessions to remember context; guardrails live on the endpoint.

Self-check

1. Users ask questions about HR policy PDFs. Which tool? (RAG tool with a knowledge base)
2. How does the agent keep context between messages? (Sessions on the endpoint)
3. Name one guardrail available on the endpoint. (Content moderation, prompt-injection, or PII protection)

Next in the series: Part 6