Oracle Database 26ai Select AI Practical Guide: Why you need to design data movement boundaries and tool execution locations before NL2SQL

Publication date: 2026-05-05 | Category: Development information

1) One-line problem definition

Key takeaway: The most common failure when attaching AI to in-house data starts not with the model, but with not first deciding where to move the data and at what layer to run the tools.

Oracle Database 26ai's Select AI not only has the ability to convert natural language into SQL. According to the official documentation, 26ai pulls NL2SQL, RAG, agent, summarization, translation, and feedback loop all inside the database. It may seem convenient on the surface, but in practice, more important questions arise. This is because you must first decide not “Can you attach an LLM?” but Which execution boundary will place queries, searches, tool calls, and memory retention for corporate data

This article is a commentary for backend developers and data platform teams reviewing in-house analytics systems, data products, operational automation, and AI-based internal tools. The scope is Oracle Select AI 26ai's structure and adoption judgment criteria. Conversely, it does not cover general generative AI introduction, LLM model performance ranking, or full data warehouse design.

2) Conclusion first

Key takeaway: Select AI 26ai is more for Teams that want to control AI execution close to the data than “attaching a chatbot to the database” That’s right.

• A team that is a good fit right now: A team that already has an Oracle database as its core system and wants to quickly add NL2SQL, RAG, and internal agents while reducing data movement
• Teams that are still overkill: Teams where most of the data sources are outside of Oracle and browser automation or external SaaS writing tools are more important
• Key judgment axis: Rather than the model itself Execution location, governance, data export minimization, SQL accuracy feedback system

My judgment is clear. The real value of Select AI 26ai is not “creating SQL in natural language,” but designto complete data query, RAG, and agent execution as much as possible within the database security boundary. Therefore, it is better suited to teams that need to add AI features while maintaining audit logs and permission models, rather than teams looking for a simple chat interface.

3) Core structure decomposition

Key summary: Select AI 26ai is not a single function, but can be understood when viewed as a system that includes Profile, NL2SQL, RAG, Agent, Feedback It’s easy.

3-1. AI Profile layer

The first thing to look at is not the model, but the AI ​​profile created with DBMS_CLOUD_AI.CREATE_PROFILE. This profile contains conditions such as which provider to use, which credentials to use, and which objects to allow for queries. To put it simply for a novice developer, this is the step of creating Data Access Agreement before the prompt.

3-2. NL2SQL layer

Select AI handles natural language queries and SQL generation and explanation with actions such as SELECT AI, SHOWSQL, EXPLAINSQL, and CHAT. The important point is that it is not only simple creation, but also and the generated SQL can be shown and explained. This structure is beneficial for teams that need to review “what the AI ​​did”

3-3. RAG layer

26ai's RAG is designed to manage the flow of chunking object storage documents, creating embeddings, and updating vector indexes on the database side without creating a separate external search pipeline. In other words, the approach is to attach RAG not as a separate product, but as auxiliary search layer in the database.

3-4. Agent layer

Select AI Agent follows the ReAct pattern and operates in four layers: Planning → Tool execution → Reflection → Memory management. Tools that can be used here include NL2SQL, RAG, custom PL/SQL, and external REST API. In other words, beyond “the database answers questions,” it becomes an execution frame that reads data, selects tools, and continues with subsequent actions.

3-5. Feedback layer

Where the Oracle documentation is particularly practical is in feedback loops. You can give positive or negative feedback on the SQL generation results with DBMS_CLOUD_AI.FEEDBACK. This feature is not just convenience,Correct NL2SQL accuracy during operation due to in-house terminology and schema specificityThis is a device that does this.

4) Explanation of design intent

Key takeaways: Oracle's direction is less about model connectivity itself and more about putting enterprise data and AI execution under the same governance

When adding AI functions, many teams subtract the data into ETL, create a separate vector database, and place the agent runtime on a separate server. This approach is flexible, but as data passes through multiple layers, permissions, recency, and audit trails are prone to lapses. Oracle 26ai took the opposite approach. The goal is to bundle natural language queries, vector searches, and agent tool execution around the database Reduce data movement and leverage existing security perimeters

It is clear what you gain and what you give up.

• What you get: Data close execution, audit and authorization model reuse, SQL/PLSQL-based fast composition, integrated operation of RAG and agent
• What you give up: Complete vendor independence, simplicity in a non-Oracle-centric stack, and freedom of browser and external SaaS-centric automation
• Practical interpretation: Although it is strong in data-driven business automation, it may be excessive when viewed as a basic platform for screen-operated agents or general-purpose multi-SaaS automation

In my interpretation, it is correct to view Select AI 26ai as a case where the data layer began to absorb part of the agent runtime, rather than “LLM was put into the DB.” Understanding this difference will ensure that your decision to adopt is not wavered.

5) Evidence and comparison

Key takeaways: Select AI 26ai should not be compared with simple NL2SQL tools, but rather with Data movement/operational boundaries/feedback possibilities

The key evidence confirmed based on official documents is as follows.

• Select AI example document: SELECT AI, SHOWSQL, EXPLAINSQL, CHAT, SUMMARIZE, FEEDBACK are provided as one grammar system.
• Supported provider range: OCI, OpenAI, Cohere, Azure OpenAI, Google, Anthropic, Hugging Face, AWS, OpenAI-compatible provider examples are provided in the official documentation. In other words, the range of model providers is wide, but the operation center is DB.
• RAG document: Specifies that documents such as PDF, DOC, JSON, XML, HTML, etc. will be converted to plain text and automate chunking, embedding, and vector index updates.
• Agent document: Specifies the four layers of Planning, Tool Use, Reflection, and Memory Management and the ReAct loop.
• Security controls: The RAG document states that administrators can disable actual table data or vector lookup documents from being sent to LLM.

In other words, when looking at Select AI 26ai, the key is not “how smart a model can be used,” but How much corporate data-based AI can be operated with less movement and less separate infrastructure.

6) Actual operation flow / step-by-step execution method

Key takeaway: Rather than uploading to Agent from the beginning, Restrict profile → SHOWSQL verification → Feedback → RAG → Agent The order is safe.

1. Narrow the allowed objects in your AI profile
   Do not open the entire schema from scratch, only the tables you actually need, for example sales.orders, sales.customers Put in object_list
2. Verify with SHOWSQL before RUNSQL.
   Before executing immediately, it is necessary to expose and review the generated SQL. In the beginning, it is better to avoid automatic execution without human inspection.
3. Create a feedback loop.
   Unusual internal terminology may sway the first results. Accuracy can be stabilized by accumulating frequently incorrect patterns to DBMS_CLOUD_AI.FEEDBACK.
4. Separates structured data and document data.
   Table queries are divided into NL2SQL, and manual/policy/document searches are divided into RAG. Cramming two problems into the same prompt reduces accuracy and makes debugging difficult.
5. RAG starts with a small set of documents.
   Put the documents into object storage and create a vector index, but first put only one domain document to check the retrieval quality.
6. Agent starts with a reading tool.
   Do not allow external REST write or change operations from the beginning, only open NL2SQL·RAG·Query REST to see the agent's plan quality first.

-- 1) Example of AI profile creation
BEGIN
  DBMS_CLOUD_AI.CREATE_PROFILE(
    profile_name => 'GENAI',
    attributes => '{
      "provider": "oci",
      "credential_name": "OCI$RESOURCE_PRINCIPAL",
      "object_list": [
        {"owner": "SH", "name": "CUSTOMERS"},
        {"owner": "SH", "name": "COUNTRIES"}
      ]
    }'
  );
END;
/

-- 2) Apply profile to session
EXEC DBMS_CLOUD_AI.SET_PROFILE('GENAI');

-- 3) Check the generated SQL first
SELECT AI SHOWSQL how many customers in San Francisco are married;

-- 4) Example of vector index creation for RAG
BEGIN
  DBMS_CLOUD_AI.CREATE_VECTOR_INDEX(
    index_name => 'RAG_INDEX',
    attributes => '{
      "vector_db_provider": "oracle",
      "profile_name": "RAG_PROFILE",
      "chunk_overlap": 128,
      "chunk_size": 1024
    }'
  );
END;
/

-- 5) Example of RAG tool creation in Agent
EXEC DBMS_CLOUD_AI_AGENT.CREATE_TOOL(
  'RAG_TOOL',
  '{"tool_type": "RAG", "tool_params": {"profile_name": "RAG_PROFILE"}}'
);

As a practical standard, you must see the two numbers below. First, SHOWSQL inspection pass rate. Second, RAG retrieved document hit rate. If these two are low and you raise Agent first, most of the automation piled on top of it will be an illusion.

7) Mistakes/Pitfalls

Key takeaway: Failure of Select AI 26ai is usually not a model problem, but rather Tolerance over-openness and hierarchy confusion It comes out.

• Mistake 1: Leaving an entire schema open in a profile
  Prevention: Minimize object_list and isolate sensitive schemas
  Recovery: Look at usage logs to reduce excessive object access. Replace with profile
• Mistake 2: Auto-execute RUNSQL immediately without SHOWSQL
  Prevention: In the initial stage, review the generated SQL and place acceptance criteria
  Recovery: Stop auto-execution and return to an explainable query path. Revert.
• Mistake 3: Treating NL2SQL and RAG as if they were the same problem
  Prevention: Separate structured queries and document retrieval into separate paths
  Recovery: Recognize failure cases as SQL mistranslations and retrieve documents. Separate to see if it fails and measure again.
• Mistake 4: Opening external writing tools on the Agent too soon
  Prevention: Separate read-only tools from approval-type operations
  Recovery: Temporarily remove REST writing tools and based on reflection logs Adjust policy.
• Mistake 5: Not using the feedback function
  Prevention: Record frequently incorrect in-house terms and join patterns and provide feedback.
  Recovery: Group incorrectly answered queries and conduct regular correction sessions. It is operated.

8) Strengths and limitations

Key summary: Select AI 26ai is strong in data-driven AI operations, but overestimating it as a general-purpose agent platform would be misleading.

Strengths

• NL2SQL, SQL description, RAG, Agent, and feedback loop within one operational boundary.
• Existing Oracle security, permissions, and auditing systems can be reused, making it easier to explain to regulated industries.
• The automation of chunking, embedding, and index update in document-based RAG is relatively clear.
• Tie together PL/SQL and REST tools to quickly create internal automation scenarios.

Limit

• The structure itself may be excessive for organizations that are not centered around Oracle.
• Separate runtimes are more natural for browser manipulation, external SaaS multi-write, and UI-centric workflows.
• NL2SQL easily creates illusions of accuracy without feedback and metadata quality.
• The presence of Agent does not replace approval boundaries and policy design.

Counterexample: If you are a team that performs sales automation or marketing operations across multiple SaaS, an external orchestration platform may be better than an agent inside Oracle DB.

9) Points to study more deeply

Key summary: The next step is not to memorize the list of functions, but to decide which tasks will be completed in the DB

• NL2SQL Accuracy: How much influence comments, foreign keys, constraints, and column descriptions have on generation quality
• RAG operation: Which document group to vectorize, and whether chunk size 1024 and overlap 128 fit our document
• Agent policy: How to separate readers and writers
• Memory strategy: Boundary between information that can be left in long-term memory and session-specific information
• Governance: When to turn on the disable LLM transport policy for real table data and vector documents

If you are a novice developer, you can get started by clearly understanding two things: “Why do you need to limit the object_list before generating SQL?” and “Why do you need to separate RAG and table queries?”

10) Execution Checklist + Author’s Perspective

Key summary: The completion standard for the introduction of Select AI 26ai is not a demo, but The acceptance range and inspection loop are alive. Status.

• Have you minimized the allowable tables and credential scope in your AI profile?
• Are you inspecting the initial query focusing on SHOWSQL rather than RUNSQL?
• Is there a feedback loop for frequently missed terms and join patterns?
• Are structured data queries and document RAG divided into different paths?
• Are external write permissions among Agent tools sent behind the approval boundary?
• Have you documented any scenarios where tables/documents should be prevented from being sent to LLM?

Definition of Done: If a limited AI profile, SHOWSQL inspection procedure, feedback record, read-oriented Agent policy, and RAG recovery quality measurement table are in place, and 30 test queries are repeated without fatal misexecution, the first introduction is considered complete.

My recommendations:If Oracle-based enterprise data is already a core asset, Select AI 26ai is a pretty strong option. However, do not start from the beginning by saying, “I will do everything from the DB to the agent.”Stabilize NL2SQL inspection loop and RAG recovery qualityIt is better to upload an agent after doing so. Conversely, if your data sources are distributed and your primary focus is external app automation, it is more realistic to limit Select AI to a secondary data layer rather than centering it.

Reference material

• Oracle Machine Learning Blog - Select AI by release: a quick guide to 26ai and 19c capabilities
• Oracle Docs - Examples of Using Select AI (Confirmed 2026-05-05)
• Oracle Docs - About Select AI Agent (Confirmed 2026-05-05)
• Oracle Docs - Select AI with Retrieval Augmented Generation (RAG) (Confirmed 2026-05-05)