The 1Z0-1159-26 exam is of medium difficulty if you already write SQL and have built a small retrieval prototype, and hard if you have never tuned a vector index. You need about 34 of 50 answers right (68%) in 90 minutes, and 56% of the marks sit in just two objectives: vector distance and indexing, and RAG.
This entry is about difficulty, the question people search as "1Z0-1159-26 難易度". Instead of one rating for the whole exam, it rates each of the six objectives, explains why some feel harder than their weight suggests, and shows how the picture changes with your background. Then come a six-week plan, a worked example, the mistakes that make the exam feel harder than it is, and five short answers. Every exam figure was read on Oracle's own pages on 2 November 2026.
What does 1Z0-1159-26 look like on paper?
1Z0-1159-26 is the exam for the Agentic AI for Oracle AI Database Certified Professional credential. It is a proctored online, multiple choice exam with 50 questions, a 90-minute limit and a 68% passing score, under Oracle's Cloud Recertification policy.
Those figures come from the official 1Z0-1159-26 exam page on Oracle MyLearn, which also lists the six objectives and their weights:
- Explain Oracle AI Vector Search Foundations: 8%
- Apply Vector Distance and Indexing: 32%
- Build Retrieval-Augmented Generation Solutions: 24%
- Design Select AI Agent Solutions: 16%
- Configure Agentic Interfaces and Tool Access: 12%
- Govern Private Agent Factory Deployments: 8%
Two numbers set the pace. At 68% you can miss about 16 questions, if every question counts the same. And 90 minutes for 50 questions gives you 1.8 minutes each, which is enough for a careful read but not for working a scenario out from scratch.
Why does 1Z0-1159-26 feel harder than a typical database exam?
It mixes three skill sets in one sitting: SQL and PL/SQL, machine learning ideas such as embeddings and similarity, and integration work such as agents, tools and MCP endpoints. Few candidates are strong in all three, so almost everyone has one weak third.
The wording of the objectives tells you more. Oracle splits the six objectives into 14 sub-objectives. Only 5 of them start with "Explain". The other 9 start with Apply, Analyze, Build, Use, Design, Configure or Choose. That means most questions are likely to describe a situation and ask what you would do, not ask you to recall a definition. Recall questions are quick. Judgment questions take the full 1.8 minutes.
Pricing is simple. The Oracle certification FAQ on exam attempts values one exam attempt at US$245, held in your MyLearn account until you register. Each attempt is a separate purchase, which is one more reason to judge your readiness topic by topic before you book.
Which 1Z0-1159-26 objectives are the hardest?
Vector distance and indexing is the hardest and also the heaviest, at 32%, about 16 questions. RAG comes second in both weight and difficulty. The two short governance and foundations objectives are the easiest marks on the paper.
The question counts below are the weight multiplied by 50. Oracle does not publish per-objective counts, and the difficulty column is our own estimate for a candidate with SQL experience.
- Vector Search Foundations, 8%, about 4 questions: easy. Concepts, embedding models, ONNX import, vectors generated inside the database.
- Vector Distance and Indexing, 32%, about 16 questions: hard. Metrics, exact versus approximate search, HNSW versus IVF, memory and accuracy trade-offs, multi-vector search.
- Retrieval-Augmented Generation, 24%, about 12 questions: medium to hard. Chunking, embeddings, retrieval, prompt building, and the PL/SQL packages that run it.
- Select AI Agent Solutions, 16%, about 8 questions: medium. Tools, tasks, agents, agent teams and how they run.
- Agentic Interfaces and Tool Access, 12%, about 6 questions: medium. The ADB MCP Server and calling Select AI Agent tools from MCP clients.
- Private Agent Factory Governance, 8%, about 4 questions: easy to medium. No-code agent use cases, roles, security controls.
Why is vector distance and indexing so demanding?
Because every question in it is a trade-off. The Oracle AI Vector Search User's Guide describes several distance metrics: Euclidean and Euclidean squared, cosine, dot product, Manhattan, Hamming and Jaccard. Its advice is to match the metric to the one the embedding model was trained with. Cosine is the default when you create a vector index without naming one.
Then come the two index families. HNSW is an in-memory neighbor graph, so memory planning matters. IVF is a neighbor partition index, and the guide warns that inserts, updates and deletes after it is built can shift the data and lower accuracy until you rebuild. Exact search compares the query with every vector and gives perfect quality at a cost in time. Approximate search uses the index and a target accuracy. The guide's own example: if a top-5 approximate search finds 4 of the 5 vectors an exact search finds, accuracy is 80%.
Small rules hide in this objective too. All vectors in an indexed column need the same number of dimensions. A column can carry only one type of vector index. And the optimizer only uses a vector index when the query's distance function is the same as the index's. Questions that hinge on one of these details are where prepared candidates still lose marks.
What makes the RAG objective tricky?
RAG questions follow the whole flow: split documents into chunks, create embeddings, find the closest chunks, and build a prompt for the language model. A wrong choice at any step shows up as a bad answer at the end, so questions often describe the symptom and ask for the cause.
Oracle's About Select AI guide adds a detail worth knowing: for SQL generation, Select AI sends the model schema metadata only, not row values, while the narrate action does pass query results or retrieved content to the model. A question about what data leaves the database turns on exactly that difference.
Are the agent and MCP objectives hard?
They are medium, because they are newer and less familiar rather than deep. The Use MCP Server guide for Autonomous AI Database shows the moving parts: you enable the server with an OCI free-form tag named adb$feature, create custom tools with DBMS_CLOUD_AI_AGENT.CREATE_TOOL, and point an MCP client that supports streamable HTTP at the database's endpoint. When you disable the server, new calls stop and calls already running finish. Read the steps once with a console open and most of these questions become recall.
How hard is 1Z0-1159-26 for your background?
Difficulty depends on which third of the exam is new to you. Find your row, then spend the extra time where it says.
- Oracle DBA: indexing will feel familiar in spirit, but distance metrics and embeddings will not. Put most of your extra time into the 8% foundations and the RAG flow.
- SQL or PL/SQL developer: RAG with PL/SQL packages will come naturally. Index memory, accuracy targets and HNSW versus IVF choices need the most practice.
- Python or machine learning engineer: embeddings, chunking and similarity will be easy. Oracle's own syntax, the packages, and the MCP and OCI tag steps are the gap.
- Architect or team lead: the design and governance objectives suit you. The 32% indexing objective needs hands-on time, because reading alone will not be enough.
Is there a Japanese version of 1Z0-1159-26?
Yes. Oracle lists 1Z0-1159-26-JPN, the Japanese-language exam page on MyLearn. It has the same 50 questions and 68% passing score and leads to the same credential. The one difference on the page is the time: 91 minutes instead of 90.
日本語での要点: 1Z0-1159-26 は全50問、合格ライン68%、英語版は90分、日本語版 (1Z0-1159-26-JPN) は91分です。当ブログの見立てでは、配点の32%を占める「ベクトル距離と索引」が最も難しい分野で、RAG の24%と合わせると56%になります。
What does a six-week study plan look like?
Six weeks at about six hours a week works for most candidates with SQL experience. Put the heavy objectives early, so you have time to go back to them.
- Week 1: foundations. Read the vector search concepts, load an ONNX embedding model, and generate vectors inside the database.
- Week 2: distance metrics and exact search. Run the same query with cosine, dot and Euclidean squared, and compare the order of results.
- Week 3: indexes. Build an HNSW index and an IVF index on the same data, change the target accuracy, and note how speed and results move.
- Week 4: RAG. Build one complete flow: chunk a document set, embed it, retrieve, and generate an answer with Select AI.
- Week 5: agents and MCP. Create one Select AI Agent tool, group tools into a task and an agent, and call it through an MCP client.
- Week 6: governance and review. Read the Private Agent Factory roles and controls, then redo week 3, your weakest area for most people.
Worked example: one support agent, five objectives
A company wants an assistant that answers staff questions from 40,000 product manuals stored in Oracle AI Database. This is not an exam item, but it shows how one project touches five of the six objectives.
First the manuals are split into chunks and embedded with a model loaded into the database (foundations). Engineers then choose cosine distance, because the embedding model was trained with it, and an HNSW index, because the vector space is large and answers must be fast (distance and indexing). Retrieved chunks are combined with the question into a prompt for the language model (RAG). A Select AI Agent gets two tools, one for search and one to open a support ticket, organised as a task (agent design). Finally the team enables the MCP Server so the company's chat client can call those tools (interfaces and tool access).
When the manuals change every day, the team has to think about index maintenance and accuracy over time. That one sentence is the kind of detail a 32% objective is built on.
Which mistakes make 1Z0-1159-26 feel harder than it is?
Most lost marks come from habits, not from missing knowledge. These five are the most common.
- Studying by objective count instead of weight. Six objectives look equal on a list, but one of them is four times the size of the smallest.
- Reading about indexes without building one. The trade-offs only stick after you have changed a target accuracy and watched results change.
- Mixing up the two Select AI data paths. Remember what metadata-only means for SQL generation and what narrate passes on.
- Treating MCP as theory. Enabling it is a tag and an endpoint; a short hands-on pass makes those questions quick.
- Spending three minutes on one scenario. At 1.8 minutes a question, mark it, move on, and come back.
Which video helps you warm up?
The DBExam channel has a walkthrough of under seven minutes on vector search and RAG ideas for this exam. Watch it after week 2, when the terms are no longer new.
Frequently asked questions
1. How many questions are on 1Z0-1159-26 and what score do I need?
The exam has 50 multiple choice questions, a 90-minute time limit and a passing score of 68%, which is about 34 correct answers if every question counts the same.
2. Which 1Z0-1159-26 topic carries the most weight?
Apply Vector Distance and Indexing, at 32%. Build Retrieval-Augmented Generation Solutions follows at 24%, so the two together make up 56% of the exam.
3. Can I take 1Z0-1159-26 in Japanese?
Yes. The Japanese exam, 1Z0-1159-26-JPN, has 50 questions, a 68% passing score and 91 minutes, and it earns the same Agentic AI for Oracle AI Database Certified Professional credential.
4. How much does the 1Z0-1159-26 exam cost?
Oracle's certification FAQ values one exam attempt at US$245. The attempt is held in your MyLearn account and used when you register for the exam.
5. Do I need hands-on practice to pass 1Z0-1159-26?
Yes. Nine of the fourteen sub-objectives ask you to apply, analyze, build, use, design, configure or choose, so practice with vectors, indexes, a RAG flow and an agent tool matters more than reading alone.