Free 1Z0-1159-26 Cheat Sheet 2026 — Topic Weights, Exam Traps & Quick Rules

Core Framework

1Z0-1159-26 Section Map

Section names come from a secondary source and Oracle publishes no weights, so cover all six evenly. Expect short SQL or PL/SQL fragments and design constraints: which package, function, index or deployment choice satisfies a privacy, memory or review requirement.

01
Vector Search Foundations (weight not published) — VECTOR type and formats, in-database ONNX vs. provider embeddings, TO_VECTOR
02
Vector Distance and Indexing (weight not published) — VECTOR_DISTANCE metrics, FETCH EXACT vs. APPROX, HNSW vs. IVF, hybrid indexes
03
RAG Solutions (weight not published) — DBMS_VECTOR_CHAIN pipeline, chunk overlap, credentials and ACLs, Select AI with RAG
04
Select AI Agent Solutions (weight not published) — AI profiles, RUNSQL vs. SHOWSQL vs. NARRATE vs. CHAT, tools, tasks, teams
05
Agentic Interfaces and Tool Access (weight not published) — MCP roles, ADB MCP Server vs. SQLcl, least privilege, VPD
06
Private Agent Factory Governance (weight not published) — self-deployed container, Agent Builder, local models, approved connections

Scenario Traps

Wrong instinct vs correct approach

Regulated text must never leave the database, yet rows need embeddings generated on insert

✕ Wrong instinct

Call a hosted provider through UTL_TO_EMBEDDING with a stored credential

✓ Correct approach

Any external provider receives the input text. Load an ONNX model with DBMS_VECTOR.LOAD_ONNX_MODEL and call VECTOR_EMBEDDING in the INSERT, so embedding runs inside the database boundary.

A RAG pipeline on an on-premises database times out calling the LLM provider

✕ Wrong instinct

Recreate the credential with DBMS_VECTOR.CREATE_CREDENTIAL

✓ Correct approach

A host or connection error is not an authorization error. Outside Autonomous AI Database, outbound hosts need network ACL entries as well as a credential. Remember UTL_TO_GENERATE_TEXT only generates; retrieval is a separate similarity query.

Analysts must review SQL that Select AI generates before anything runs against production

✕ Wrong instinct

Use the default RUNSQL action and log the queries afterwards

✓ Correct approach

SHOWSQL returns the generated SQL without executing it. NARRATE summarises query results and CHAT skips the database entirely. Narrow the profile object list and add comments or annotations rather than renaming cryptic columns.

Support users search manuals by exact part numbers and error codes, and pure vector search misses them

✕ Wrong instinct

Increase chunk size and re-embed the manuals

✓ Correct approach

Embeddings blur exact tokens. A hybrid vector index combines Oracle Text keywords with similarity, queried through DBMS_HYBRID_VECTOR.SEARCH, so rare identifiers rank correctly. Bigger chunks mix topics and weaken similarity further.

A third-party assistant needs payroll data via MCP, with strict compliance on sensitive columns

✕ Wrong instinct

Sign in as ADMIN and add a system prompt forbidding the pay-rate columns

✓ Correct approach

Prompts are guidance, not enforcement. Use the managed ADB MCP Server with a least-privilege user, focused tools rather than one run-any-SQL tool, and VPD or redaction, with unified auditing recording tool-driven queries.

Quick Rules

Know these cold

▸ Formats — LOAT32 balanced, INT8 smaller, BINARY smallest and compared with HAMMING or JACCARD
▸ COSINE is the default metric and ignores magnitude; EUCLIDEAN is sensitive to magnitude
▸ FETCH EXACT returns true top-k at full cost; FETCH APPROX uses TARGET ACCURACY and an index
▸ HNSW NEIGHBORS and EFCONSTRUCTION raise accuracy and build cost; they are not IVF settings
▸ Pipeline order — TL_TO_TEXT, UTL_TO_CHUNKS with overlap, UTL_TO_EMBEDDINGS, retrieve, UTL_TO_GENERATE_TEXT
▸ Use PARTITIONS BY in multi-vector search so one document cannot flood the top results
▸ RAG adds current data at query time and can cite sources; fine-tuning changes behaviour slowly
▸ For SQL generation, Select AI sends schema metadata from the object list, not table rows
▸ MCP servers advertise tools; clients only discover and call what the server exposes
▸ Private Agent Factory is a container you run; local Ollama or vLLM meets no-egress rules

Self Check

Can you answer these without checking your notes?

In this scenario: "Regulated text must never leave the database, yet rows need embeddings generated on insert" — what should you do first?

Any external provider receives the input text. Load an ONNX model with DBMS_VECTOR.LOAD_ONNX_MODEL and call VECTOR_EMBEDDING in the INSERT, so embedding runs inside the database boundary.

In this scenario: "A RAG pipeline on an on-premises database times out calling the LLM provider" — what should you do first?

A host or connection error is not an authorization error. Outside Autonomous AI Database, outbound hosts need network ACL entries as well as a credential. Remember UTL_TO_GENERATE_TEXT only generates; retrieval is a separate similarity query.

In this scenario: "Analysts must review SQL that Select AI generates before anything runs against production" — what should you do first?

SHOWSQL returns the generated SQL without executing it. NARRATE summarises query results and CHAT skips the database entirely. Narrow the profile object list and add comments or annotations rather than renaming cryptic columns.

Failure Patterns

Common Exam Mistakes — What candidates get wrong

Sorting VECTOR_DISTANCE results in descending order to get the best matches

Oracle returns distances, so smaller means more similar and the best rows come first in ascending order. DOT is the negated dot product for the same reason. VECTOR_NORM returns one vector's magnitude and never compares two items.

Querying with a metric that differs from the vector index DISTANCE clause

The optimizer uses a vector index only when the query metric matches the index's DISTANCE. FETCH APPROX without a matching index silently falls back to an exact scan, so adding the clause alone gives no speed gain.

Choosing HNSW for a corpus too large for memory

HNSW graphs live in the vector pool sized by VECTOR_MEMORY_SIZE. IVF, built as NEIGHBOR PARTITIONS around k-means centroids, is disk-based and scales past memory, so enlarging the vector pool never fixes an IVF index.

Creating agent objects with DBMS_CLOUD_AI instead of DBMS_CLOUD_AI_AGENT

DBMS_CLOUD_AI holds profiles and SET_PROFILE for natural language to SQL. Tools, agents, tasks and teams live in DBMS_CLOUD_AI_AGENT, tool access is granted on the task, and a team must be activated with SET_TEAM before SELECT AI AGENT.

Comparing vectors from two embedding models because both have 768 dimensions

Each model defines its own vector space, so query and stored vectors must come from one model, and a model change means re-embedding the corpus. Declare the VECTOR column with dimensions matching the model's output.

Continue your prep

Resources to help you pass Agentic AI for Oracle AI Database on your first attempt

Practice Questions

Strengthen weak areas with exam-style practice questions and detailed explanations.

Mock Exam

Simulate the real exam experience and assess your readiness under timed conditions.

Study Guide

Review key concepts, objectives, and exam topics in one place.

AI Tutor

Get personalized explanations, learning recommendations, and instant answers.

1Z0-1159-26 Training

Follow a structured learning path designed to help you prepare efficiently.