The agents require reliable information and controlled operations. Oracle AI Database provides close to data capabilities: vector search, agent construction, orchestration with Select AI Agent and integration with MCP. Each resolves a different part of the problem.
Key concepts
AI Vector Search finds semantically close elements. Private Agent Factory offers visual construction of agents and workflows. Select AI Agent It allows to configure agents, tasks, tools and equipment using database capabilities. The MCP Server from Autonomous AI Database Expand tools for supported customers. One does not automatically replace the other.
Module Objectives
- Explain embeddings, search for similarity and its difference with RAG.
- Track production, storage, indexation and vector consultation.
- Choose from search, visual construction, agent equipment and MCP exposure.
- Keep permissions data even if the consultation arrives in a natural language.
Vector and similarity
A embedding represents data using a list of numbers generated by a model. Their usefulness depends on how relevant relationships are reflected in that space. Two phrases can be about the same without sharing exact words, and a semantic search can get that relationship back.
For text, the path is text → tokenizer → model → vector. tokens can be words or parts of words: To count words doesn't mean to count tokens. If the entry limit of the model is exceeded, content can be truncated. That's why we divide long documents before we generate embeddings. A fair size and overlap between fragments to preserve context and check recovery with representative questions without applying a universal number. Oracle explain these changes and Fragmentation parameters.
The document and consultation embeddings model should produce supported representations: change dimensions or model without regenerating and validating data breaks comparison. The distance metric also matters. Cosine focuses on guidance. Euclidean measures geometric separation. Choose the recommended metric for the model and check its conduct with examples of your domain.
The column VECTOR allows to store embeddings with identifiers, text, prices or permissions. Vector search can be combined with relational filters. "Semantically similar" means neither "true" nor "authorized"; apply access restrictions before providing model document.
Vector Search Workflow
- Prepare and divide content into understandable fragments, preserving origin and metadata.
- Generates embeddings with a compatible model, inside the base with ONNX as appropriate or with an outside service.
- Save your vectors with your business data and your identifiers.
- Assess whether a vector index should be created: HNSW using an e-graph structure IVF grouping vectors into partitions. A index accelerates some scenarios with resources commitments and recovery.
- Generates the embedding of the consultation, looks for similarity and applying SQL filters. If you build RAG, give the relevant fragments to LLM to formulate a response with sources.
That tour relies on the official workflow of AI Vector Search. Indexing is optional by scale and target. an exact search can be useful as a reference. RAG adds generation about recovered content: this isn't another name to save vectors.
In an implementation, DBMS_VECTOR_CHAIN allows to Chain UTL_TO_TEXT to extract text, UTL_TO_CHUNKS to divide and UTL_TO_EMBEDDINGS To produce vectors: the output of each stage feeds the following. You have to set up the model and check the requirements of Oracle Text and the chosen supplier. and utilities guide details these conditions.
Step-by-step example
A tent has backpack descriptions and user wants "a light backpack to wear laptop, for less than 200 soles."
First, it retains description, price, category and embedding of each product. Convert demand into a vector compatible with existing ones. It recovers semantically close candidates and filters price and availability with structured data. It orders and verifies that the description really mentions space for laptop. Finally, it presents options and their attributes supported by the table.
That educational SQL means a table productoshis column embedding of type VECTOR and the reference vector linked as :query_embedding. Use COSINE only if the model corresponds to them. FETCH EXACT makes explicit an exact search for the three closest neighbours that fulfill the filters. Query exact official search syntax.
SELECT nombre, precio, descripcion
FROM productos
WHERE precio < 200 AND stock > 0
ORDER BY VECTOR_DISTANCE(embedding, :query_embedding, COSINE)
FETCH EXACT FIRST 3 ROWS ONLY;
Also try an unmatching consultation and a product that's semantically close but outweighs the price. The right answer may be "I did not find an option that met both conditions." The similarity does not allow them to ignore restrictions and to invent stock.
A complete solution: answer with evidence
The two branches solve different needs. The semantic search Reforms relevant fragments to substantiate an answer. a SQL consultation gets structured results, for example the state and an order amount. Both require monitoring what information can be found with the conduct of an operation. To change data or run a business action also requires an authorized tool and its validation.
How to read the image: for "what policy does it apply?" select documentary recovery; for "how much did this client pay?" indicates structured consultation. Connectors represent reference and response relationships, and they do not represent a rigid sequence. No receives model unrestricted direct access to the base, nor does this view mean that LLM runs within Oracle AI Database.
Private Agent Factory
Agent Factory offers prepared agents and a visual builder to connect sources, models and tools. Special work was undertaken in the course's notes Knowledge Agentand Data Analysis Agent, focused on structured data. The documentation for version 26.7 also includes Deep Data Research Agent; prevents memorizing a number of agents as if they never change. Query official introduction.
For an assistant policy, set up an approved source, check ingestion, ask questions with a known answer and check citations. For analysis, check allowed tables and column meaning. The visual builder facilitates assembly but still requires access monitoring, testing and deployment. "Private" does not show that all models are housed on your network: check endpoints and datashow.
Reproduces first the minimum circuit: Chat Input → Agent with a LLM configured → Chat Output. Connects compatible inputs and inputs; Without an additional source or tool, that flow doesn't check your business data. The official components explain these connections. Save and Test at Playground; then publish the flow and configure the authentication of the endpoint to invoke it from another application, following the Agent Builder guide.
Select AI Agent
Select AI Agent articula tools, tasks, agents y Teams. A tool offers an operation A task defines jobs. an agent does so with his or her capabilities; A team coordinates inputs. There may be a sequential order or a supervisor that select tasks. History views help to observe executions, according to official concepts.
The package DBMS_CLOUD_AI_AGENT manage these capabilities and AI profiles configure the model's access. That the agent is integrated with the base does not mean that the weights of the LLM are housed inside the engine. Don't confuse that case with running an embeddings model ONNX admitted into the base.
MCP at Autonomous AI Database
The managed server allows you to discover and invoke Select AI Agent tools from supported MCP applications. Identification and database controls are still important: credentials, roles, row policies, network and audit delimits access. A profile with frame context helps to generate queries but does not guarantee SQL right or replace permissions. The official architecture of the server describes these layers.
Distinguish three tools implementations: one tool SQL uses an NL2SQL profile to translate questions. A tool RAG regains knowledge by means of a vector search. A custom tool calls a PL/SQL functionfor instance to consult inventory. They are registered with DBMS_CLOUD_AI_AGENT.CREATE_TOOLand his types and requirements. Generate SQL does not authorize your execution: the application and tool have to limit the actions that have been authorized and the base applying the privileges of the performing identity. Check what tools are registered by MCP.
Common errors
- Consider an embedding a legible copy of the document.
- To assume that an approximate index always recovers exactly the same neighbours as an exact search.
- To merge generation of SQL with permission to run.
- To say that using a local model removes all latency.
Selection exercise
You want two outside applications to consult an authorized inventory feature without hosting another MCP server. What's his capacity and what aren't you supposed to give up?
Solution
The MCP Server managed Autonomous AI Database can display Select AI Agent framework tools as supported version and configuration. Still you have to define the tool, permissions, client's identity and connectivity. The protocol does not guarantee that each user can consult any inventory and does not convert a reading function into a write authorization.
Sources and review
Go back to AI Vector Search, Agent Factory, Select AI Agent y Autonomous MCP Server. To close the module, assign a capacity to every need: search, assemble, orchestrate or display tools.